Merge b610945825 into d0f0b15cf5

Update ComfyUI screenshot to showcase a more modern workflow

README.md

@@ -1,7 +1,7 @@
 <div align="center">
 
 # ComfyUI
-**The most powerful and modular visual AI engine and application.**
+**The most powerful and modular AI engine for content creation.**
 
 
 [![Website][website-shield]][website-url]
@@ -31,10 +31,16 @@
 [github-downloads-latest-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/latest/total?style=flat&label=downloads%40latest
 [github-downloads-link]: https://github.com/comfyanonymous/ComfyUI/releases
 
-![ComfyUI Screenshot](https://github.com/user-attachments/assets/7ccaf2c1-9b72-41ae-9a89-5688c94b7abe)
+<img width="1590" height="795" alt="ComfyUI Screenshot" src="https://github.com/user-attachments/assets/36e065e0-bfae-4456-8c7f-8369d5ea48a2" />
+<br>
 </div>
 
-ComfyUI lets you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. Available on Windows, Linux, and macOS.
+ComfyUI is the AI creation engine for visual professionals who demand control over every model, every parameter, and every output. Its powerful and modular node graph interface empowers creatives to generate images, videos, 3D models, audio, and more...
+- ComfyUI natively supports the latest open-source state of the art models.
+- API nodes provide access to the best closed source models such as Nano Banana, Seedance, Hunyuan3D, etc.
+- It is available on Windows, Linux, and macOS, locally with our desktop application or on our cloud.
+- The most sophisticated workflows can be exposed through a simple UI thanks to App Mode.
+- It integrates seamlessly into production pipelines with our API endpoints.
 
 ## Get Started
 
@@ -77,6 +83,7 @@ See what ComfyUI can do with the [newer template workflows](https://comfy.org/wo
    - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/)
    - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/)
    - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/)
+   - Ernie Image
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)

comfy/ldm/lightricks/av_model.py

@@ -16,6 +16,7 @@ from comfy.ldm.lightricks.model import (
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
 from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import comfy.ldm.common_dit
+import comfy.model_prefetch
 
 class CompressedTimestep:
     """Store video timestep embeddings in compressed form using per-frame indexing."""
@@ -907,9 +908,11 @@ class LTXAVModel(LTXVModel):
         """Process transformer blocks for LTXAV."""
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        prefetch_queue = comfy.model_prefetch.make_prefetch_queue(list(self.transformer_blocks), vx.device, transformer_options)
 
         # Process transformer blocks
         for i, block in enumerate(self.transformer_blocks):
+            comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, block)
             if ("double_block", i) in blocks_replace:
 
                 def block_wrap(args):
@@ -982,6 +985,8 @@ class LTXAVModel(LTXVModel):
                     a_prompt_timestep=a_prompt_timestep,
                 )
 
+        comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, None)
+
         return [vx, ax]
 
     def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):

comfy/ldm/modules/attention.py

@@ -14,6 +14,8 @@ from .sub_quadratic_attention import efficient_dot_product_attention
 
 from comfy import model_management
 
+TORCH_HAS_GQA = model_management.torch_version_numeric >= (2, 5)
+
 if model_management.xformers_enabled():
     import xformers
     import xformers.ops
@@ -150,7 +152,12 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         b, _, dim_head = q.shape
         dim_head //= heads
 
-    scale = dim_head ** -0.5
+    if kwargs.get("enable_gqa", False) and q.shape[-3] != k.shape[-3]:
+        n_rep = q.shape[-3] // k.shape[-3]
+        k = k.repeat_interleave(n_rep, dim=-3)
+        v = v.repeat_interleave(n_rep, dim=-3)
+
+    scale = kwargs.get("scale", dim_head ** -0.5)
 
     h = heads
     if skip_reshape:
@@ -219,6 +226,10 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
         b, _, dim_head = query.shape
         dim_head //= heads
 
+    if "scale" in kwargs:
+        # Pre-scale query to match requested scale (cancels internal 1/sqrt(dim_head))
+        query = query * (kwargs["scale"] * dim_head ** 0.5)
+
     if skip_reshape:
         query = query.reshape(b * heads, -1, dim_head)
         value = value.reshape(b * heads, -1, dim_head)
@@ -290,7 +301,7 @@ def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         b, _, dim_head = q.shape
         dim_head //= heads
 
-    scale = dim_head ** -0.5
+    scale = kwargs.get("scale", dim_head ** -0.5)
 
     if skip_reshape:
          q, k, v = map(
@@ -500,8 +511,13 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
         if mask.ndim == 3:
             mask = mask.unsqueeze(1)
 
+    # Pass through extra SDPA kwargs (scale, enable_gqa) if provided
+    # enable_gqa requires PyTorch 2.5+; older versions use manual KV expansion above
+    sdpa_keys = ("scale", "enable_gqa") if TORCH_HAS_GQA else ("scale",)
+    sdpa_extra = {k: v for k, v in kwargs.items() if k in sdpa_keys}
+
     if SDP_BATCH_LIMIT >= b:
-        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False, **sdpa_extra)
         if not skip_output_reshape:
             out = (
                 out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@@ -519,7 +535,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
                 k[i : i + SDP_BATCH_LIMIT],
                 v[i : i + SDP_BATCH_LIMIT],
                 attn_mask=m,
-                dropout_p=0.0, is_causal=False
+                dropout_p=0.0, is_causal=False, **sdpa_extra
             ).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head)
     return out
 

comfy/lora.py

@@ -17,6 +17,7 @@
 """
 
 from __future__ import annotations
+import comfy.memory_management
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
@@ -473,3 +474,17 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             weight = old_weight
 
     return weight
+
+def prefetch_prepared_value(value, allocate_buffer, stream):
+    if isinstance(value, torch.Tensor):
+        dest = allocate_buffer(comfy.memory_management.vram_aligned_size(value))
+        comfy.model_management.cast_to_gathered([value], dest, non_blocking=True, stream=stream)
+        return comfy.memory_management.interpret_gathered_like([value], dest)[0]
+    elif isinstance(value, weight_adapter.WeightAdapterBase):
+        return type(value)(value.loaded_keys, prefetch_prepared_value(value.weights, allocate_buffer, stream))
+    elif isinstance(value, tuple):
+        return tuple(prefetch_prepared_value(item, allocate_buffer, stream) for item in value)
+    elif isinstance(value, list):
+        return [prefetch_prepared_value(item, allocate_buffer, stream) for item in value]
+
+    return value

comfy/model_base.py

@@ -214,6 +214,11 @@ class BaseModel(torch.nn.Module):
         if "latent_shapes" in extra_conds:
             xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes"))
 
+        transformer_options = transformer_options.copy()
+        transformer_options["prefetch_dynamic_vbars"] = (
+            self.current_patcher is not None and self.current_patcher.is_dynamic()
+        )
+
         model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
         if len(model_output) > 1 and not torch.is_tensor(model_output):
             model_output, _ = utils.pack_latents(model_output)

comfy/model_management.py

@@ -31,6 +31,7 @@ from contextlib import nullcontext
 import comfy.memory_management
 import comfy.utils
 import comfy.quant_ops
+import comfy_aimdo.vram_buffer
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -1175,6 +1176,10 @@ stream_counters = {}
 
 STREAM_CAST_BUFFERS = {}
 LARGEST_CASTED_WEIGHT = (None, 0)
+STREAM_AIMDO_CAST_BUFFERS = {}
+LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+
+DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3
 
 def get_cast_buffer(offload_stream, device, size, ref):
     global LARGEST_CASTED_WEIGHT
@@ -1208,13 +1213,26 @@ def get_cast_buffer(offload_stream, device, size, ref):
 
     return cast_buffer
 
+def get_aimdo_cast_buffer(offload_stream, device):
+    cast_buffer = STREAM_AIMDO_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None:
+        cast_buffer = comfy_aimdo.vram_buffer.VRAMBuffer(DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE, device.index)
+        STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer
+
+    return cast_buffer
 def reset_cast_buffers():
     global LARGEST_CASTED_WEIGHT
+    global LARGEST_AIMDO_CASTED_WEIGHT
+
     LARGEST_CASTED_WEIGHT = (None, 0)
-    for offload_stream in STREAM_CAST_BUFFERS:
-        offload_stream.synchronize()
+    LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+    for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS):
+        if offload_stream is not None:
+            offload_stream.synchronize()
     synchronize()
+
     STREAM_CAST_BUFFERS.clear()
+    STREAM_AIMDO_CAST_BUFFERS.clear()
     soft_empty_cache()
 
 def get_offload_stream(device):

comfy/model_patcher.py

@@ -121,9 +121,20 @@ class LowVramPatch:
         self.patches = patches
         self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
+        self.prepared_patches = None
+
+    def prepare(self, allocate_buffer, stream):
+        self.prepared_patches = [
+            (patch[0], comfy.lora.prefetch_prepared_value(patch[1], allocate_buffer, stream), patch[2], patch[3], patch[4])
+            for patch in self.patches[self.key]
+        ]
+
+    def clear_prepared(self):
+        self.prepared_patches = None
 
     def __call__(self, weight):
-        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+        patches = self.prepared_patches if self.prepared_patches is not None else self.patches[self.key]
+        return comfy.lora.calculate_weight(patches, weight, self.key, intermediate_dtype=weight.dtype)
 
 LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
 

comfy/model_prefetch.py

@@ -0,0 +1,65 @@
+import comfy_aimdo.model_vbar
+import comfy.model_management
+import comfy.ops
+
+PREFETCH_QUEUES = []
+
+def cleanup_prefetched_modules(comfy_modules):
+    for s in comfy_modules:
+        prefetch = getattr(s, "_prefetch", None)
+        if prefetch is None:
+            continue
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                lowvram_fn.clear_prepared()
+        if prefetch["signature"] is not None:
+            comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        delattr(s, "_prefetch")
+
+def cleanup_prefetch_queues():
+    global PREFETCH_QUEUES
+
+    for queue in PREFETCH_QUEUES:
+        for entry in queue:
+            if entry is None or not isinstance(entry, tuple):
+                continue
+            _, prefetch_state = entry
+            comfy_modules = prefetch_state[1]
+            if comfy_modules is not None:
+                cleanup_prefetched_modules(comfy_modules)
+    PREFETCH_QUEUES = []
+
+def prefetch_queue_pop(queue, device, module):
+    if queue is None:
+        return
+
+    consumed = queue.pop(0)
+    if consumed is not None:
+        offload_stream, prefetch_state = consumed
+        offload_stream.wait_stream(comfy.model_management.current_stream(device))
+        _, comfy_modules = prefetch_state
+        if comfy_modules is not None:
+            cleanup_prefetched_modules(comfy_modules)
+
+    prefetch = queue[0]
+    if prefetch is not None:
+        comfy_modules = []
+        for s in prefetch.modules():
+            if hasattr(s, "_v"):
+                comfy_modules.append(s)
+
+        offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True)
+        comfy.model_management.sync_stream(device, offload_stream)
+        queue[0] = (offload_stream, (prefetch, comfy_modules))
+
+def make_prefetch_queue(queue, device, transformer_options):
+    if (not transformer_options.get("prefetch_dynamic_vbars", False)
+        or comfy.model_management.NUM_STREAMS == 0
+        or comfy.model_management.is_device_cpu(device)
+        or not comfy.model_management.device_supports_non_blocking(device)):
+        return None
+
+    queue = [None] + queue + [None]
+    PREFETCH_QUEUES.append(queue)
+    return queue

comfy/ops.py

@@ -86,38 +86,61 @@ def materialize_meta_param(s, param_keys):
             setattr(s, param_key, torch.nn.Parameter(torch.zeros(param.shape, dtype=param.dtype), requires_grad=param.requires_grad))
 
 
-def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
-    #vbar doesn't support CPU weights, but some custom nodes have weird paths
-    #that might switch the layer to the CPU and expect it to work. We have to take
-    #a clone conservatively as we are mmapped and some SFT files are packed misaligned
-    #If you are a custom node author reading this, please move your layer to the GPU
-    #or declare your ModelPatcher as CPU in the first place.
-    if comfy.model_management.is_device_cpu(device):
-        materialize_meta_param(s, ["weight", "bias"])
-        weight = s.weight.to(dtype=dtype, copy=True)
-        if isinstance(weight, QuantizedTensor):
-            weight = weight.dequantize()
-        bias = None
-        if s.bias is not None:
-            bias = s.bias.to(dtype=bias_dtype, copy=True)
-        return weight, bias, (None, None, None)
-
+# FIXME: add n=1 cache hit fast path
+def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blocking):
     offload_stream = None
-    xfer_dest = None
+    cast_buffer = None
+    cast_buffer_offset = 0
+
+    def ensure_offload_stream(module, required_size, check_largest):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+
+        if offload_stream is None:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+        if offload_stream is None or not check_largest or len(comfy_modules) != 1:
+            return
+
+        current_size = 0 if cast_buffer is None else cast_buffer.size()
+        if current_size < required_size and module is comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[0]:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+            cast_buffer = None
+        if required_size > comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[1]:
+            comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT = (module, required_size)
+
+    def get_cast_buffer(buffer_size):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+        nonlocal cast_buffer_offset
+
+        if buffer_size == 0:
+            return None
+
+        if offload_stream is None:
+            return torch.empty((buffer_size,), dtype=torch.uint8, device=device)
+
+        cast_buffer = comfy.model_management.get_aimdo_cast_buffer(offload_stream, device)
+        buffer = comfy_aimdo.torch.aimdo_to_tensor(cast_buffer.get(buffer_size, cast_buffer_offset), device)
+        cast_buffer_offset += buffer_size
+        return buffer
+
+    for s in comfy_modules:
+        signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
+        resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+        prefetch = {
+            "signature": signature,
+            "resident": resident,
+        }
 
-    signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
-    resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
-    if signature is not None:
         if resident:
-            weight = s._v_weight
-            bias = s._v_bias
-        else:
-            xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+            s._prefetch = prefetch
+            continue
 
-    if not resident:
         materialize_meta_param(s, ["weight", "bias"])
+        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) if signature is not None else None
         cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
         cast_dest = None
+        needs_cast = False
 
         xfer_source = [ s.weight, s.bias ]
 
@@ -129,22 +152,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             if data is None:
                 continue
             if data.dtype != geometry.dtype:
+                needs_cast = True
                 cast_dest = xfer_dest
-                if cast_dest is None:
-                    cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
                 xfer_dest = None
                 break
 
         dest_size = comfy.memory_management.vram_aligned_size(xfer_source)
-        offload_stream = comfy.model_management.get_offload_stream(device)
-        if xfer_dest is None and offload_stream is not None:
-                xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
-                if xfer_dest is None:
-                    offload_stream = comfy.model_management.get_offload_stream(device)
-                    xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+        ensure_offload_stream(s, dest_size if xfer_dest is None else 0, True)
         if xfer_dest is None:
-            xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device)
-            offload_stream = None
+            xfer_dest = get_cast_buffer(dest_size)
 
         if signature is None and pin is None:
             comfy.pinned_memory.pin_memory(s)
@@ -157,27 +173,54 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             xfer_source = [ pin ]
         #send it over
         comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream)
-        comfy.model_management.sync_stream(device, offload_stream)
 
-        if cast_dest is not None:
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                ensure_offload_stream(s, cast_buffer_offset, False)
+                lowvram_fn.prepare(lambda size: get_cast_buffer(size), offload_stream)
+
+        prefetch["xfer_dest"] = xfer_dest
+        prefetch["cast_dest"] = cast_dest
+        prefetch["cast_geometry"] = cast_geometry
+        prefetch["needs_cast"] = needs_cast
+        s._prefetch = prefetch
+
+    return offload_stream
+
+
+def resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant):
+
+    prefetch = getattr(s, "_prefetch", None)
+
+    if prefetch["resident"]:
+        weight = s._v_weight
+        bias = s._v_bias
+    else:
+        xfer_dest = prefetch["xfer_dest"]
+        if prefetch["needs_cast"]:
+            cast_dest = prefetch["cast_dest"] if prefetch["cast_dest"] is not None else torch.empty((comfy.memory_management.vram_aligned_size(prefetch["cast_geometry"]),), dtype=torch.uint8, device=device)
             for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest),
-                                           comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)):
+                                           comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], cast_dest)):
                 if post_cast is not None:
                     post_cast.copy_(pre_cast)
             xfer_dest = cast_dest
 
-        params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
+        params = comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], xfer_dest)
         weight = params[0]
         bias = params[1]
-        if signature is not None:
+        if prefetch["signature"] is not None:
             s._v_weight = weight
             s._v_bias = bias
-        s._v_signature=signature
+        s._v_signature = prefetch["signature"]
 
     def post_cast(s, param_key, x, dtype, resident, update_weight):
         lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
         fns = getattr(s, param_key + "_function", [])
 
+        if x is None:
+            return None
+
         orig = x
 
         def to_dequant(tensor, dtype):
@@ -205,14 +248,12 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             x = f(x)
         return x
 
-    update_weight = signature is not None
+    update_weight = prefetch["signature"] is not None
+    weight = post_cast(s, "weight", weight, dtype, prefetch["resident"], update_weight)
+    if bias is not None:
+        bias = post_cast(s, "bias", bias, bias_dtype, prefetch["resident"], update_weight)
 
-    weight = post_cast(s, "weight", weight, dtype, resident, update_weight)
-    if s.bias is not None:
-        bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight)
-
-    #FIXME: weird offload return protocol
-    return weight, bias, (offload_stream, device if signature is not None else None, None)
+    return weight, bias
 
 
 def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None, want_requant=False):
@@ -230,10 +271,46 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         if device is None:
             device = input.device
 
+    def format_return(result, offloadable):
+        weight, bias, offload_stream = result
+        return (weight, bias, offload_stream) if offloadable else (weight, bias)
+
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
 
     if hasattr(s, "_v"):
-        return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant)
+
+        #vbar doesn't support CPU weights, but some custom nodes have weird paths
+        #that might switch the layer to the CPU and expect it to work. We have to take
+        #a clone conservatively as we are mmapped and some SFT files are packed misaligned
+        #If you are a custom node author reading this, please move your layer to the GPU
+        #or declare your ModelPatcher as CPU in the first place.
+        if comfy.model_management.is_device_cpu(device):
+            materialize_meta_param(s, ["weight", "bias"])
+            weight = s.weight.to(dtype=dtype, copy=True)
+            if isinstance(weight, QuantizedTensor):
+                weight = weight.dequantize()
+            bias = s.bias.to(dtype=bias_dtype, copy=True) if s.bias is not None else None
+            return format_return((weight, bias, (None, None, None)), offloadable)
+
+        prefetched = hasattr(s, "_prefetch")
+        offload_stream = None
+        offload_device = None
+        if not prefetched:
+            offload_stream = cast_modules_with_vbar([s], dtype, device, bias_dtype, non_blocking)
+            comfy.model_management.sync_stream(device, offload_stream)
+
+        weight, bias = resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant)
+
+        if not prefetched:
+            if getattr(s, "_prefetch")["signature"] is not None:
+                offload_device = device
+            for param_key in ("weight", "bias"):
+                lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+                if lowvram_fn is not None:
+                    lowvram_fn.clear_prepared()
+            delattr(s, "_prefetch")
+        return format_return((weight, bias, (offload_stream, offload_device, None)), offloadable)
+
 
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
@@ -280,11 +357,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         for f in s.weight_function:
             weight = f(weight)
 
-    if offloadable:
-        return weight, bias, (offload_stream, weight_a, bias_a)
-    else:
-        #Legacy function signature
-        return weight, bias
+    return format_return((weight, bias, (offload_stream, weight_a, bias_a)), offloadable)
 
 
 def uncast_bias_weight(s, weight, bias, offload_stream):
@@ -1173,6 +1246,93 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                         self._buffers[key] = fn(buf)
                 return self
 
+        class Embedding(manual_cast.Embedding):
+            def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                    strict, missing_keys, unexpected_keys, error_msgs):
+                weight_key = f"{prefix}weight"
+                layer_conf = state_dict.pop(f"{prefix}comfy_quant", None)
+                if layer_conf is not None:
+                    layer_conf = json.loads(layer_conf.numpy().tobytes())
+
+                # Only fp8 makes sense for embeddings (per-row dequant via index select).
+                # Block-scaled formats (NVFP4, MXFP8) can't do per-row lookup efficiently.
+                quant_format = layer_conf.get("format", None) if layer_conf is not None else None
+                if quant_format in ["float8_e4m3fn", "float8_e5m2"] and weight_key in state_dict:
+                    self.quant_format = quant_format
+                    qconfig = QUANT_ALGOS[quant_format]
+                    layout_cls = get_layout_class(qconfig["comfy_tensor_layout"])
+                    weight = state_dict.pop(weight_key)
+                    manually_loaded_keys = [weight_key]
+
+                    scale_key = f"{prefix}weight_scale"
+                    scale = state_dict.pop(scale_key, None)
+                    if scale is not None:
+                        scale = scale.float()
+                        manually_loaded_keys.append(scale_key)
+
+                    params = layout_cls.Params(
+                        scale=scale if scale is not None else torch.ones((), dtype=torch.float32),
+                        orig_dtype=MixedPrecisionOps._compute_dtype,
+                        orig_shape=(self.num_embeddings, self.embedding_dim),
+                    )
+                    self.weight = torch.nn.Parameter(
+                        QuantizedTensor(weight.to(dtype=qconfig["storage_t"]), qconfig["comfy_tensor_layout"], params),
+                        requires_grad=False)
+
+                    super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+                    for k in manually_loaded_keys:
+                        if k in missing_keys:
+                            missing_keys.remove(k)
+                else:
+                    if layer_conf is not None:
+                        state_dict[f"{prefix}comfy_quant"] = torch.tensor(list(json.dumps(layer_conf).encode('utf-8')), dtype=torch.uint8)
+                    super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+
+            def state_dict(self, *args, destination=None, prefix="", **kwargs):
+                if destination is not None:
+                    sd = destination
+                else:
+                    sd = {}
+
+                if not hasattr(self, 'weight') or self.weight is None:
+                    return sd
+
+                if isinstance(self.weight, QuantizedTensor):
+                    sd_out = self.weight.state_dict("{}weight".format(prefix))
+                    for k in sd_out:
+                        sd[k] = sd_out[k]
+
+                    quant_conf = {"format": self.quant_format}
+                    sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
+                else:
+                    sd["{}weight".format(prefix)] = self.weight
+                return sd
+
+            def forward_comfy_cast_weights(self, input, out_dtype=None):
+                weight = self.weight
+
+                # Optimized path: lookup in fp8, dequantize only the selected rows.
+                if isinstance(weight, QuantizedTensor) and len(self.weight_function) == 0:
+                    qdata, _, offload_stream = cast_bias_weight(self, device=input.device, dtype=weight.dtype, offloadable=True)
+                    if isinstance(qdata, QuantizedTensor):
+                        scale = qdata._params.scale
+                        qdata = qdata._qdata
+                    else:
+                        scale = None
+
+                    x = torch.nn.functional.embedding(
+                        input, qdata, self.padding_idx, self.max_norm,
+                        self.norm_type, self.scale_grad_by_freq, self.sparse)
+                    uncast_bias_weight(self, qdata, None, offload_stream)
+                    target_dtype = out_dtype if out_dtype is not None else weight._params.orig_dtype
+                    x = x.to(dtype=target_dtype)
+                    if scale is not None and scale != 1.0:
+                        x = x * scale.to(dtype=target_dtype)
+                    return x
+
+                # Fallback for non-quantized or weight_function (LoRA) case
+                return super().forward_comfy_cast_weights(input, out_dtype=out_dtype)
+
     return MixedPrecisionOps
 
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):

comfy/rmsnorm.py

@@ -3,6 +3,7 @@ import comfy.model_management
 
 RMSNorm = torch.nn.RMSNorm
 
+# Note: torch's fused F.rms_norm is faster but produces slightly different output than manual implementations (rsqrt/reduction rounding).
 def rms_norm(x, weight=None, eps=1e-6):
     if weight is None:
         return torch.nn.functional.rms_norm(x, (x.shape[-1],), eps=eps)

comfy/sd.py

@@ -65,6 +65,7 @@ import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
 import comfy.text_encoders.qwen35
 import comfy.text_encoders.ernie
+import comfy.text_encoders.gemma4
 
 import comfy.model_patcher
 import comfy.lora
@@ -1271,6 +1272,9 @@ class TEModel(Enum):
     QWEN35_9B = 26
     QWEN35_27B = 27
     MINISTRAL_3_3B = 28
+    GEMMA_4_E4B = 29
+    GEMMA_4_E2B = 30
+    GEMMA_4_31B = 31
 
 
 def detect_te_model(sd):
@@ -1296,6 +1300,12 @@ def detect_te_model(sd):
             return TEModel.BYT5_SMALL_GLYPH
         return TEModel.T5_BASE
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
+        if 'model.layers.59.self_attn.q_norm.weight' in sd:
+            return TEModel.GEMMA_4_31B
+        if 'model.layers.41.self_attn.q_norm.weight' in sd and 'model.layers.47.self_attn.q_norm.weight' not in sd:
+            return TEModel.GEMMA_4_E4B
+        if 'model.layers.34.self_attn.q_norm.weight' in sd and 'model.layers.41.self_attn.q_norm.weight' not in sd:
+            return TEModel.GEMMA_4_E2B
         if 'model.layers.47.self_attn.q_norm.weight' in sd:
             return TEModel.GEMMA_3_12B
         if 'model.layers.0.self_attn.q_norm.weight' in sd:
@@ -1435,6 +1445,13 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             else:
                 clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
                 clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
+        elif te_model in (TEModel.GEMMA_4_E4B, TEModel.GEMMA_4_E2B, TEModel.GEMMA_4_31B):
+            variant = {TEModel.GEMMA_4_E4B: comfy.text_encoders.gemma4.Gemma4_E4B,
+                       TEModel.GEMMA_4_E2B: comfy.text_encoders.gemma4.Gemma4_E2B,
+                       TEModel.GEMMA_4_31B: comfy.text_encoders.gemma4.Gemma4_31B}[te_model]
+            clip_target.clip = comfy.text_encoders.gemma4.gemma4_te(**llama_detect(clip_data), model_class=variant)
+            clip_target.tokenizer = variant.tokenizer
+            tokenizer_data["tokenizer_json"] = clip_data[0].get("tokenizer_json", None)
         elif te_model == TEModel.GEMMA_2_2B:
             clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer

comfy/text_encoders/llama.py

@@ -521,7 +521,7 @@ class Attention(nn.Module):
             else:
                 present_key_value = (xk, xv, index + num_tokens)
 
-            if sliding_window is not None and xk.shape[2] > sliding_window:
+            if sliding_window is not None and xk.shape[2] > sliding_window and seq_length == 1:
                 xk = xk[:, :, -sliding_window:]
                 xv = xv[:, :, -sliding_window:]
                 attention_mask = attention_mask[..., -sliding_window:] if attention_mask is not None else None
@@ -533,12 +533,12 @@ class Attention(nn.Module):
         return self.o_proj(output), present_key_value
 
 class MLP(nn.Module):
-    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
+    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None, intermediate_size=None):
         super().__init__()
-        ops = ops or nn
-        self.gate_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
-        self.up_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
-        self.down_proj = ops.Linear(config.intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype)
+        intermediate_size = intermediate_size or config.intermediate_size
+        self.gate_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype)
+        self.up_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype)
+        self.down_proj = ops.Linear(intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype)
         if config.mlp_activation == "silu":
             self.activation = torch.nn.functional.silu
         elif config.mlp_activation == "gelu_pytorch_tanh":
@@ -647,24 +647,25 @@ class TransformerBlockGemma2(nn.Module):
 
         return x, present_key_value
 
+def _make_scaled_embedding(ops, vocab_size, hidden_size, scale, device, dtype):
+    class ScaledEmbedding(ops.Embedding):
+        def forward(self, input_ids, out_dtype=None):
+            return super().forward(input_ids, out_dtype=out_dtype) * scale
+    return ScaledEmbedding(vocab_size, hidden_size, device=device, dtype=dtype)
+
+
 class Llama2_(nn.Module):
     def __init__(self, config, device=None, dtype=None, ops=None):
         super().__init__()
         self.config = config
         self.vocab_size = config.vocab_size
 
-        self.embed_tokens = ops.Embedding(
-            config.vocab_size,
-            config.hidden_size,
-            device=device,
-            dtype=dtype
-        )
         if self.config.transformer_type == "gemma2" or self.config.transformer_type == "gemma3":
             transformer = TransformerBlockGemma2
-            self.normalize_in = True
+            self.embed_tokens = _make_scaled_embedding(ops, config.vocab_size, config.hidden_size, config.hidden_size ** 0.5, device, dtype)
         else:
             transformer = TransformerBlock
-            self.normalize_in = False
+            self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype)
 
         self.layers = nn.ModuleList([
             transformer(config, index=i, device=device, dtype=dtype, ops=ops)
@@ -690,15 +691,12 @@ class Llama2_(nn.Module):
                                     self.config.rope_dims,
                                     device=device)
 
-    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None, input_ids=None):
         if embeds is not None:
             x = embeds
         else:
             x = self.embed_tokens(x, out_dtype=dtype)
 
-        if self.normalize_in:
-            x *= self.config.hidden_size ** 0.5
-
         seq_len = x.shape[1]
         past_len = 0
         if past_key_values is not None and len(past_key_values) > 0:
@@ -850,7 +848,7 @@ class BaseGenerate:
                                     torch.empty([batch, model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype), 0))
         return past_key_values
 
-    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0):
+    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0, initial_input_ids=None):
         device = embeds.device
 
         if stop_tokens is None:
@@ -875,14 +873,16 @@ class BaseGenerate:
         pbar = comfy.utils.ProgressBar(max_length)
 
         # Generation loop
+        current_input_ids = initial_input_ids
         for step in tqdm(range(max_length), desc="Generating tokens"):
-            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values)
+            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values, input_ids=current_input_ids)
             logits = self.logits(x)[:, -1]
             next_token = self.sample_token(logits, temperature, top_k, top_p, min_p, repetition_penalty, initial_tokens + generated_token_ids, generator, do_sample=do_sample, presence_penalty=presence_penalty)
             token_id = next_token[0].item()
             generated_token_ids.append(token_id)
 
             embeds = self.model.embed_tokens(next_token).to(execution_dtype)
+            current_input_ids = next_token if initial_input_ids is not None else None
             pbar.update(1)
 
             if token_id in stop_tokens:

comfy/text_encoders/lt.py

@@ -93,8 +93,7 @@ class Gemma3_12BModel(sd1_clip.SDClipModel):
 
     def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, presence_penalty):
         tokens_only = [[t[0] for t in b] for b in tokens]
-        embeds, _, _, embeds_info = self.process_tokens(tokens_only, self.execution_device)
-        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        embeds, _, _, _ = self.process_tokens(tokens_only, self.execution_device)
         return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[106], presence_penalty=presence_penalty)  # 106 is <end_of_turn>
 
 class DualLinearProjection(torch.nn.Module):

comfy/text_encoders/lumina2.py

@@ -50,8 +50,7 @@ class Gemma3_4B_Vision_Model(sd1_clip.SDClipModel):
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B_Vision, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
     def process_tokens(self, tokens, device):
-        embeds, _, _, embeds_info = super().process_tokens(tokens, device)
-        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        embeds, _, _, _ = super().process_tokens(tokens, device)
         return embeds
 
 class LuminaModel(sd1_clip.SD1ClipModel):

comfy/text_encoders/qwen35.py

@@ -408,8 +408,6 @@ class Qwen35Transformer(Llama2_):
         nn.Module.__init__(self)
         self.config = config
         self.vocab_size = config.vocab_size
-        self.normalize_in = False
-
         self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype)
         self.layers = nn.ModuleList([
             Qwen35TransformerBlock(config, index=i, device=device, dtype=dtype, ops=ops)

comfy/utils.py

@@ -1446,10 +1446,3 @@ def deepcopy_list_dict(obj, memo=None):
     memo[obj_id] = res
     return res
 
-def normalize_image_embeddings(embeds, embeds_info, scale_factor):
-    """Normalize image embeddings to match text embedding scale"""
-    for info in embeds_info:
-        if info.get("type") == "image":
-            start_idx = info["index"]
-            end_idx = start_idx + info["size"]
-            embeds[:, start_idx:end_idx, :] /= scale_factor

comfy_api/latest/_io.py

@@ -9,6 +9,7 @@ from dataclasses import asdict, dataclass, field
 from enum import Enum
 from typing import Any, Callable, Literal, TypedDict, TypeVar, TYPE_CHECKING
 from typing_extensions import NotRequired, final
+from urllib.parse import urlparse
 
 # used for type hinting
 import torch
@@ -43,7 +44,67 @@ class UploadType(str, Enum):
     model = "file_upload"
 
 
+class RemoteItemSchema:
+    """Describes how to map API response objects to rich dropdown items.
+
+    All *_field parameters use dot-path notation (e.g. ``"labels.gender"``).
+    ``label_field`` and ``description_field`` additionally support template strings
+    with ``{field}`` placeholders (e.g. ``"{name} ({labels.accent})"``).
+    """
+    def __init__(
+        self,
+        value_field: str,
+        label_field: str,
+        preview_url_field: str | None = None,
+        preview_type: Literal["image", "video", "audio"] = "image",
+        description_field: str | None = None,
+        search_fields: list[str] | None = None,
+    ):
+        if preview_type not in ("image", "video", "audio"):
+            raise ValueError(
+                f"RemoteItemSchema: 'preview_type' must be 'image', 'video', or 'audio'; got {preview_type!r}."
+            )
+        if search_fields is not None:
+            for f in search_fields:
+                if "{" in f or "}" in f:
+                    raise ValueError(
+                        f"RemoteItemSchema: 'search_fields' must be dot-paths, not template strings (got {f!r})."
+                    )
+        self.value_field = value_field
+        """Dot-path to the unique identifier within each item.
+        This value is stored in the widget and passed to execute()."""
+        self.label_field = label_field
+        """Dot-path to the display name, or a template string with {field} placeholders."""
+        self.preview_url_field = preview_url_field
+        """Dot-path to a preview media URL. If None, no preview is shown."""
+        self.preview_type = preview_type
+        """How to render the preview: "image", "video", or "audio"."""
+        self.description_field = description_field
+        """Optional dot-path or template for a subtitle line shown below the label."""
+        self.search_fields = search_fields
+        """Dot-paths to fields included in the search index. When unset, search falls back to
+        the resolved label (i.e. ``label_field`` after template substitution). Note that template
+        label strings (e.g. ``"{first} {last}"``) are not valid path entries here — list the
+        underlying paths (``["first", "last"]``) instead."""
+
+    def as_dict(self):
+        return prune_dict({
+            "value_field": self.value_field,
+            "label_field": self.label_field,
+            "preview_url_field": self.preview_url_field,
+            "preview_type": self.preview_type,
+            "description_field": self.description_field,
+            "search_fields": self.search_fields,
+        })
+
+
 class RemoteOptions:
+    """Plain remote combo: fetches a list of strings/objects and populates a standard dropdown.
+
+    Use this for lightweight lists from endpoints that return a bare array (or an array under
+    ``response_key``). For rich dropdowns with previews, search, filtering, or pagination,
+    use :class:`RemoteComboOptions` and the ``remote_combo=`` parameter on ``Combo.Input``.
+    """
     def __init__(self, route: str, refresh_button: bool, control_after_refresh: Literal["first", "last"]="first",
                  timeout: int=None, max_retries: int=None, refresh: int=None):
         self.route = route
@@ -70,6 +131,113 @@ class RemoteOptions:
         })
 
 
+class RemoteComboOptions:
+    """Rich remote combo: populates a Vue dropdown with previews, search, filtering, and pagination.
+
+    Attached to a :class:`Combo.Input` via ``remote_combo=`` (not ``remote=``). Requires an
+    ``item_schema`` describing how to map API response objects to dropdown items.
+
+    Response-shape contract:
+    - Without ``page_size``: endpoint returns an array (or an array at ``response_key``).
+    - With ``page_size``: endpoint returns ``{"items": [...], "has_more": bool}`` and is fetched
+      progressively, appending each page to the dropdown.
+
+    Pagination contract (when ``page_size`` is set):
+    - The frontend issues ``GET <route>?page=<n>&page_size=<size>`` with ``page`` starting at ``0``
+      and incrementing by 1 until the endpoint returns ``has_more: false`` or an empty ``items`` list.
+    - Endpoints that use 1-based pages, ``limit``/``offset``, or cursor/continuation tokens are not
+      supported directly - adapt them via the proxy or
+      expose a small shim endpoint that translates to the ``page`` + ``page_size`` + ``{items, has_more}`` shape.
+    """
+    def __init__(
+        self,
+        route: str,
+        item_schema: RemoteItemSchema,
+        refresh_button: bool = True,
+        auto_select: Literal["first", "last"] | None = None,
+        timeout: int | None = None,
+        max_retries: int | None = None,
+        refresh: int | None = None,
+        response_key: str | None = None,
+        use_comfy_api: bool = False,
+        page_size: int | None = None,
+    ):
+        if page_size is not None:
+            if response_key is not None:
+                raise ValueError(
+                    "RemoteComboOptions: pass 'response_key' or 'page_size', not both. "
+                    "Paginated responses must use the top-level 'items' field."
+                )
+            if page_size < 1:
+                raise ValueError(
+                    f"RemoteComboOptions: 'page_size' must be >= 1 when set (got {page_size})."
+                )
+        if auto_select is not None and auto_select not in ("first", "last"):
+            raise ValueError(
+                f"RemoteComboOptions: 'auto_select' must be 'first', 'last', or None; got {auto_select!r}."
+            )
+        if refresh is not None and 0 < refresh < 128:
+            raise ValueError(
+                f"RemoteComboOptions: 'refresh' must be >= 128 (ms TTL) or <= 0 (cache never expires); got {refresh}."
+            )
+        if timeout is not None and timeout < 0:
+            raise ValueError(
+                f"RemoteComboOptions: 'timeout' must be >= 0 (got {timeout})."
+            )
+        if max_retries is not None and max_retries < 0:
+            raise ValueError(
+                f"RemoteComboOptions: 'max_retries' must be >= 0 (got {max_retries})."
+            )
+        if not route.startswith("/"):
+            parsed = urlparse(route)
+            if not (parsed.scheme and parsed.netloc):
+                raise ValueError(
+                    f"RemoteComboOptions: 'route' must start with '/' or be an absolute URL; got {route!r}."
+                )
+            if use_comfy_api:
+                raise ValueError(
+                    f"RemoteComboOptions: 'use_comfy_api=True' cannot be combined with absolute URL {route!r}."
+                )
+        self.route = route
+        """The route to the remote source."""
+        self.item_schema = item_schema
+        """Required: describes how each API response object maps to a dropdown item."""
+        self.refresh_button = refresh_button
+        """Specifies whether to show a refresh button next to the widget."""
+        self.auto_select = auto_select
+        """Fallback item to select when the widget's value is empty. Never overrides an existing
+        selection. Default None means no fallback."""
+        self.timeout = timeout
+        """Maximum time to wait for a response, in milliseconds."""
+        self.max_retries = max_retries
+        """Maximum number of retries before aborting the request. Default None uses the frontend's built-in limit."""
+        self.refresh = refresh
+        """TTL of the cached value in milliseconds. Must be >= 128 (ms TTL) or <= 0 (cache never expires,
+        re-fetched only via the refresh button). Default None uses the frontend's built-in behavior."""
+        self.response_key = response_key
+        """Dot-path to the items array in a non-paginated response. Mutually exclusive with
+        ``page_size``; paginated responses must use the top-level ``items`` field."""
+        self.use_comfy_api = use_comfy_api
+        """When True, the frontend prepends the comfy-api base URL to ``route`` and injects auth headers."""
+        self.page_size = page_size
+        """When set, switches the widget to progressive-fetch mode. The endpoint must return
+        ``{"items": [...], "has_more": bool}``."""
+
+    def as_dict(self):
+        return prune_dict({
+            "route": self.route,
+            "item_schema": self.item_schema.as_dict(),
+            "refresh_button": self.refresh_button,
+            "auto_select": self.auto_select,
+            "timeout": self.timeout,
+            "max_retries": self.max_retries,
+            "refresh": self.refresh,
+            "response_key": self.response_key,
+            "use_comfy_api": self.use_comfy_api,
+            "page_size": self.page_size,
+        })
+
+
 class NumberDisplay(str, Enum):
     number = "number"
     slider = "slider"
@@ -359,11 +527,16 @@ class Combo(ComfyTypeIO):
             upload: UploadType=None,
             image_folder: FolderType=None,
             remote: RemoteOptions=None,
+            remote_combo: RemoteComboOptions=None,
             socketless: bool=None,
             extra_dict=None,
             raw_link: bool=None,
             advanced: bool=None,
         ):
+            if remote is not None and remote_combo is not None:
+                raise ValueError("Combo.Input: pass either 'remote' or 'remote_combo', not both.")
+            if options is not None and remote_combo is not None:
+                raise ValueError("Combo.Input: pass either 'options' or 'remote_combo', not both.")
             if isinstance(options, type) and issubclass(options, Enum):
                 options = [v.value for v in options]
             if isinstance(default, Enum):
@@ -375,6 +548,7 @@ class Combo(ComfyTypeIO):
             self.upload = upload
             self.image_folder = image_folder
             self.remote = remote
+            self.remote_combo = remote_combo
             self.default: str
 
         def as_dict(self):
@@ -385,6 +559,7 @@ class Combo(ComfyTypeIO):
                 **({self.upload.value: True} if self.upload is not None else {}),
                 "image_folder": self.image_folder.value if self.image_folder else None,
                 "remote": self.remote.as_dict() if self.remote else None,
+                "remote_combo": self.remote_combo.as_dict() if self.remote_combo else None,
             })
 
     class Output(Output):
@@ -2221,7 +2396,9 @@ class NodeReplace:
 __all__ = [
     "FolderType",
     "UploadType",
+    "RemoteItemSchema",
     "RemoteOptions",
+    "RemoteComboOptions",
     "NumberDisplay",
     "ControlAfterGenerate",
 

comfy_api_nodes/apis/topaz.py

@@ -1,4 +1,4 @@
-from typing import Optional, Union
+from typing import Optional
 
 from pydantic import BaseModel, Field
 
@@ -72,8 +72,11 @@ class VideoEnhancementFilter(BaseModel):
     grain: Optional[float] = Field(None, description="Grain after AI model processing")
     grainSize: Optional[float] = Field(None, description="Size of generated grain")
     recoverOriginalDetailValue: Optional[float] = Field(None, description="Source details into the output video")
-    creativity: Optional[str] = Field(None, description="Creativity level(high, low) for slc-1 only")
+    creativity: float | str | None = Field(None, description="slc-1/slp-2.5: enum (low/middle/high). ast-2: decimal 0.0-1.0.")
     isOptimizedMode: Optional[bool] = Field(None, description="Set to true for Starlight Creative (slc-1) only")
+    prompt: str | None = Field(None, description="Descriptive scene prompt (ast-2 only)")
+    sharp: float | None = Field(None, description="ast-2 pre-enhance sharpness")
+    realism: float | None = Field(None, description="ast-2 realism control")
 
 
 class OutputInformationVideo(BaseModel):
@@ -90,7 +93,7 @@ class Overrides(BaseModel):
 
 class CreateVideoRequest(BaseModel):
     source: CreateVideoRequestSource = Field(...)
-    filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...)
+    filters: list[VideoFrameInterpolationFilter | VideoEnhancementFilter] = Field(...)
     output: OutputInformationVideo = Field(...)
     overrides: Overrides = Field(Overrides(isPaidDiffusion=True))
 

comfy_api_nodes/nodes_elevenlabs.py

@@ -233,6 +233,85 @@ class ElevenLabsVoiceSelector(IO.ComfyNode):
         return IO.NodeOutput(voice_id)
 
 
+class ElevenLabsRichVoiceSelector(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsRichVoiceSelector",
+            display_name="ElevenLabs Voice Selector (Rich)",
+            category="api node/audio/ElevenLabs",
+            description="Select an ElevenLabs voice with audio preview and rich metadata.",
+            inputs=[
+                IO.Combo.Input(
+                    "voice",
+                    remote_combo=IO.RemoteComboOptions(
+                        route="/proxy/elevenlabs/v2/voices",
+                        refresh_button=True,
+                        refresh=43200000,
+                        use_comfy_api=True,
+                        page_size=100,
+                        item_schema=IO.RemoteItemSchema(
+                            value_field="voice_id",
+                            label_field="name",
+                            preview_url_field="preview_url",
+                            preview_type="audio",
+                            search_fields=["name", "labels.gender", "labels.accent", "labels.use_case"],
+                        ),
+                    ),
+                    tooltip="Choose a voice with audio preview.",
+                ),
+            ],
+            outputs=[
+                IO.Custom(ELEVENLABS_VOICE).Output(display_name="voice"),
+            ],
+            is_api_node=False,
+        )
+
+    @classmethod
+    def execute(cls, voice: str) -> IO.NodeOutput:
+        return IO.NodeOutput(voice)  # voice is already the voice_id from item_schema.value_field
+
+
+class ElevenLabsSharedVoiceSelector(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsSharedVoiceSelector",
+            display_name="ElevenLabs Shared Voice Selector",
+            category="api node/audio/ElevenLabs",
+            description="Browse the ElevenLabs shared voice library (11K+ community voices) with audio preview.",
+            inputs=[
+                IO.Combo.Input(
+                    "voice",
+                    remote_combo=IO.RemoteComboOptions(
+                        route="/proxy/elevenlabs/v1/shared-voices",
+                        refresh_button=True,
+                        refresh=43200000,
+                        use_comfy_api=True,
+                        page_size=100,
+                        item_schema=IO.RemoteItemSchema(
+                            value_field="voice_id",
+                            label_field="name",
+                            preview_url_field="preview_url",
+                            preview_type="audio",
+                            description_field="descriptive",
+                            search_fields=["name", "gender", "accent", "use_case", "descriptive"],
+                        ),
+                    ),
+                    tooltip="Browse shared voices with audio preview.",
+                ),
+            ],
+            outputs=[
+                IO.Custom(ELEVENLABS_VOICE).Output(display_name="voice"),
+            ],
+            is_api_node=False,
+        )
+
+    @classmethod
+    def execute(cls, voice: str) -> IO.NodeOutput:
+        return IO.NodeOutput(voice)
+
+
 class ElevenLabsTextToSpeech(IO.ComfyNode):
     @classmethod
     def define_schema(cls) -> IO.Schema:
@@ -911,6 +990,8 @@ class ElevenLabsExtension(ComfyExtension):
         return [
             ElevenLabsSpeechToText,
             ElevenLabsVoiceSelector,
+            ElevenLabsRichVoiceSelector,
+            ElevenLabsSharedVoiceSelector,
             ElevenLabsTextToSpeech,
             ElevenLabsAudioIsolation,
             ElevenLabsTextToSoundEffects,

comfy_api_nodes/nodes_kling.py

@@ -3288,6 +3288,54 @@ class KlingAvatarNode(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url))
 
 
+KLING_ELEMENT_ID = "KLING_ELEMENT_ID"
+
+
+class KlingElementSelector(IO.ComfyNode):
+    """Select a Kling preset element (character, scene, effect, etc.) for use in video generation."""
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="KlingElementSelector",
+            display_name="Kling Element Selector",
+            category="api node/video/Kling",
+            description="Browse and select a Kling preset element with image preview. Elements provide consistent characters, scenes, costumes, and effects for video generation.",
+            inputs=[
+                IO.Combo.Input(
+                    "element",
+                    remote_combo=IO.RemoteComboOptions(
+                        route="/proxy/kling/v1/general/advanced-presets-elements",
+                        refresh_button=True,
+                        refresh=43200000,
+                        use_comfy_api=True,
+                        response_key="data",
+                        item_schema=IO.RemoteItemSchema(
+                            value_field="task_result.elements.0.element_id",
+                            label_field="task_result.elements.0.element_name",
+                            preview_url_field="task_result.elements.0.element_image_list.frontal_image",
+                            preview_type="image",
+                            description_field="task_result.elements.0.element_description",
+                            search_fields=["task_result.elements.0.element_name", "task_result.elements.0.element_description"],
+                        ),
+                    ),
+                    tooltip="Select a preset element to use in video generation.",
+                ),
+            ],
+            outputs=[IO.Custom(KLING_ELEMENT_ID).Output(display_name="element_id")],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=False,
+        )
+
+    @classmethod
+    async def execute(cls, element: str) -> IO.NodeOutput:
+        return IO.NodeOutput(element)
+
+
 class KlingExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -3317,6 +3365,7 @@ class KlingExtension(ComfyExtension):
             KlingVideoNode,
             KlingFirstLastFrameNode,
             KlingAvatarNode,
+            KlingElementSelector,
         ]
 
 

comfy_api_nodes/nodes_topaz.py

@@ -36,11 +36,15 @@ from comfy_api_nodes.util import (
 )
 
 UPSCALER_MODELS_MAP = {
+    "Astra 2": "ast-2",
     "Starlight (Astra) Fast": "slf-1",
     "Starlight (Astra) Creative": "slc-1",
     "Starlight Precise 2.5": "slp-2.5",
 }
 
+AST2_MAX_FRAMES = 9000
+AST2_MAX_FRAMES_WITH_PROMPT = 450
+
 
 class TopazImageEnhance(IO.ComfyNode):
     @classmethod
@@ -230,13 +234,20 @@ class TopazVideoEnhance(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TopazVideoEnhance",
-            display_name="Topaz Video Enhance",
+            display_name="Topaz Video Enhance (Legacy)",
             category="api node/video/Topaz",
             description="Breathe new life into video with powerful upscaling and recovery technology.",
             inputs=[
                 IO.Video.Input("video"),
                 IO.Boolean.Input("upscaler_enabled", default=True),
-                IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())),
+                IO.Combo.Input(
+                    "upscaler_model",
+                    options=[
+                        "Starlight (Astra) Fast",
+                        "Starlight (Astra) Creative",
+                        "Starlight Precise 2.5",
+                    ],
+                ),
                 IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
                 IO.Combo.Input(
                     "upscaler_creativity",
@@ -304,6 +315,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            is_deprecated=True,
         )
 
     @classmethod
@@ -457,12 +469,357 @@ class TopazVideoEnhance(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(final_response.download.url))
 
 
+class TopazVideoEnhanceV2(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="TopazVideoEnhanceV2",
+            display_name="Topaz Video Enhance",
+            category="api node/video/Topaz",
+            description="Breathe new life into video with powerful upscaling and recovery technology.",
+            inputs=[
+                IO.Video.Input("video"),
+                IO.DynamicCombo.Input(
+                    "upscaler_model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "Astra 2",
+                            [
+                                IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
+                                IO.Float.Input(
+                                    "creativity",
+                                    default=0.5,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Creative strength of the upscale.",
+                                ),
+                                IO.String.Input(
+                                    "prompt",
+                                    multiline=True,
+                                    default="",
+                                    tooltip="Optional descriptive (not instructive) scene prompt."
+                                    f"Capping input at {AST2_MAX_FRAMES_WITH_PROMPT} frames (~15s @ 30fps) when set.",
+                                ),
+                                IO.Float.Input(
+                                    "sharp",
+                                    default=0.5,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Pre-enhance sharpness: "
+                                    "0.0=Gaussian blur, 0.5=passthrough (default), 1.0=USM sharpening.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "realism",
+                                    default=0.0,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Pulls output toward photographic realism."
+                                    "Leave at 0 for the model default.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight (Astra) Fast",
+                            [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight (Astra) Creative",
+                            [
+                                IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
+                                IO.Combo.Input(
+                                    "creativity",
+                                    options=["low", "middle", "high"],
+                                    default="low",
+                                    tooltip="Creative strength of the upscale.",
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight Precise 2.5",
+                            [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"])],
+                        ),
+                        IO.DynamicCombo.Option("Disabled", []),
+                    ],
+                ),
+                IO.DynamicCombo.Input(
+                    "interpolation_model",
+                    options=[
+                        IO.DynamicCombo.Option("Disabled", []),
+                        IO.DynamicCombo.Option(
+                            "apo-8",
+                            [
+                                IO.Int.Input(
+                                    "interpolation_frame_rate",
+                                    default=60,
+                                    min=15,
+                                    max=240,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Output frame rate.",
+                                ),
+                                IO.Int.Input(
+                                    "interpolation_slowmo",
+                                    default=1,
+                                    min=1,
+                                    max=16,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Slow-motion factor applied to the input video. "
+                                    "For example, 2 makes the output twice as slow and doubles the duration.",
+                                    advanced=True,
+                                ),
+                                IO.Boolean.Input(
+                                    "interpolation_duplicate",
+                                    default=False,
+                                    tooltip="Analyze the input for duplicate frames and remove them.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "interpolation_duplicate_threshold",
+                                    default=0.01,
+                                    min=0.001,
+                                    max=0.1,
+                                    step=0.001,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Detection sensitivity for duplicate frames.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                    ],
+                ),
+                IO.Combo.Input(
+                    "dynamic_compression_level",
+                    options=["Low", "Mid", "High"],
+                    default="Low",
+                    tooltip="CQP level.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=[
+                    "upscaler_model",
+                    "upscaler_model.upscaler_resolution",
+                    "interpolation_model",
+                ]),
+                expr="""
+                (
+                  $model := $lookup(widgets, "upscaler_model");
+                  $res := $lookup(widgets, "upscaler_model.upscaler_resolution");
+                  $interp := $lookup(widgets, "interpolation_model");
+                  $is4k := $contains($res, "4k");
+                  $hasInterp := $interp != "disabled";
+                  $rates := {
+                    "starlight (astra) fast":     {"hd": 0.43, "uhd": 0.85},
+                    "starlight precise 2.5":      {"hd": 0.70, "uhd": 1.54},
+                    "astra 2":                    {"hd": 1.72, "uhd": 2.85},
+                    "starlight (astra) creative": {"hd": 2.25, "uhd": 3.99}
+                  };
+                  $surcharge := $is4k ? 0.28 : 0.14;
+                  $entry := $lookup($rates, $model);
+                  $base := $is4k ? $entry.uhd : $entry.hd;
+                  $hi := $base + ($hasInterp ? $surcharge : 0);
+                  $model = "disabled"
+                    ? {"type":"text","text":"Interpolation only"}
+                    : ($hasInterp
+                        ? {"type":"text","text":"~" & $string($base) & "–" & $string($hi) & " credits/src frame"}
+                        : {"type":"text","text":"~" & $string($base) & " credits/src frame"})
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        upscaler_model: dict,
+        interpolation_model: dict,
+        dynamic_compression_level: str = "Low",
+    ) -> IO.NodeOutput:
+        upscaler_choice = upscaler_model["upscaler_model"]
+        interpolation_choice = interpolation_model["interpolation_model"]
+        if upscaler_choice == "Disabled" and interpolation_choice == "Disabled":
+            raise ValueError("There is nothing to do: both upscaling and interpolation are disabled.")
+        validate_container_format_is_mp4(video)
+        src_width, src_height = video.get_dimensions()
+        src_frame_rate = int(video.get_frame_rate())
+        duration_sec = video.get_duration()
+        src_video_stream = video.get_stream_source()
+        target_width = src_width
+        target_height = src_height
+        target_frame_rate = src_frame_rate
+        filters = []
+        if upscaler_choice != "Disabled":
+            if "1080p" in upscaler_model["upscaler_resolution"]:
+                target_pixel_p = 1080
+                max_long_side = 1920
+            else:
+                target_pixel_p = 2160
+                max_long_side = 3840
+            ar = src_width / src_height
+            if src_width >= src_height:
+                # Landscape or Square; Attempt to set height to target (e.g., 2160), calculate width
+                target_height = target_pixel_p
+                target_width = int(target_height * ar)
+                # Check if width exceeds standard bounds (for ultra-wide e.g., 21:9 ARs)
+                if target_width > max_long_side:
+                    target_width = max_long_side
+                    target_height = int(target_width / ar)
+            else:
+                # Portrait; Attempt to set width to target (e.g., 2160), calculate height
+                target_width = target_pixel_p
+                target_height = int(target_width / ar)
+                # Check if height exceeds standard bounds
+                if target_height > max_long_side:
+                    target_height = max_long_side
+                    target_width = int(target_height * ar)
+            if target_width % 2 != 0:
+                target_width += 1
+            if target_height % 2 != 0:
+                target_height += 1
+            model_id = UPSCALER_MODELS_MAP[upscaler_choice]
+            if model_id == "slc-1":
+                filters.append(
+                    VideoEnhancementFilter(
+                        model=model_id,
+                        creativity=upscaler_model["creativity"],
+                        isOptimizedMode=True,
+                    )
+                )
+            elif model_id == "ast-2":
+                n_frames = video.get_frame_count()
+                ast2_prompt = (upscaler_model["prompt"] or "").strip()
+                if ast2_prompt and n_frames > AST2_MAX_FRAMES_WITH_PROMPT:
+                    raise ValueError(
+                        f"Astra 2 with a prompt is limited to {AST2_MAX_FRAMES_WITH_PROMPT} input frames "
+                        f"(~15s @ 30fps); video has {n_frames}. Clear the prompt or shorten the clip."
+                    )
+                if n_frames > AST2_MAX_FRAMES:
+                    raise ValueError(f"Astra 2 is limited to {AST2_MAX_FRAMES} input frames; video has {n_frames}.")
+                realism = upscaler_model["realism"]
+                filters.append(
+                    VideoEnhancementFilter(
+                        model=model_id,
+                        creativity=upscaler_model["creativity"],
+                        prompt=(ast2_prompt or None),
+                        sharp=upscaler_model["sharp"],
+                        realism=(realism if realism > 0 else None),
+                    )
+                )
+            else:
+                filters.append(VideoEnhancementFilter(model=model_id))
+        if interpolation_choice != "Disabled":
+            target_frame_rate = interpolation_model["interpolation_frame_rate"]
+            filters.append(
+                VideoFrameInterpolationFilter(
+                    model=interpolation_choice,
+                    slowmo=interpolation_model["interpolation_slowmo"],
+                    fps=interpolation_model["interpolation_frame_rate"],
+                    duplicate=interpolation_model["interpolation_duplicate"],
+                    duplicate_threshold=interpolation_model["interpolation_duplicate_threshold"],
+                ),
+            )
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/topaz/video/", method="POST"),
+            response_model=CreateVideoResponse,
+            data=CreateVideoRequest(
+                source=CreateVideoRequestSource(
+                    container="mp4",
+                    size=get_fs_object_size(src_video_stream),
+                    duration=int(duration_sec),
+                    frameCount=video.get_frame_count(),
+                    frameRate=src_frame_rate,
+                    resolution=Resolution(width=src_width, height=src_height),
+                ),
+                filters=filters,
+                output=OutputInformationVideo(
+                    resolution=Resolution(width=target_width, height=target_height),
+                    frameRate=target_frame_rate,
+                    audioCodec="AAC",
+                    audioTransfer="Copy",
+                    dynamicCompressionLevel=dynamic_compression_level,
+                ),
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        upload_res = await sync_op(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/topaz/video/{initial_res.requestId}/accept",
+                method="PATCH",
+            ),
+            response_model=VideoAcceptResponse,
+            wait_label="Preparing upload",
+            final_label_on_success="Upload started",
+        )
+        if len(upload_res.urls) > 1:
+            raise NotImplementedError(
+                "Large files are not currently supported. Please open an issue in the ComfyUI repository."
+            )
+        async with aiohttp.ClientSession(headers={"Content-Type": "video/mp4"}) as session:
+            if isinstance(src_video_stream, BytesIO):
+                src_video_stream.seek(0)
+                async with session.put(upload_res.urls[0], data=src_video_stream, raise_for_status=True) as res:
+                    upload_etag = res.headers["Etag"]
+            else:
+                with builtins.open(src_video_stream, "rb") as video_file:
+                    async with session.put(upload_res.urls[0], data=video_file, raise_for_status=True) as res:
+                        upload_etag = res.headers["Etag"]
+        await sync_op(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/topaz/video/{initial_res.requestId}/complete-upload",
+                method="PATCH",
+            ),
+            response_model=VideoCompleteUploadResponse,
+            data=VideoCompleteUploadRequest(
+                uploadResults=[
+                    VideoCompleteUploadRequestPart(
+                        partNum=1,
+                        eTag=upload_etag,
+                    ),
+                ],
+            ),
+            wait_label="Finalizing upload",
+            final_label_on_success="Upload completed",
+        )
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/topaz/video/{initial_res.requestId}/status"),
+            response_model=VideoStatusResponse,
+            status_extractor=lambda x: x.status,
+            progress_extractor=lambda x: getattr(x, "progress", 0),
+            price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None),
+            poll_interval=10.0,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.download.url))
+
+
 class TopazExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
         return [
             TopazImageEnhance,
             TopazVideoEnhance,
+            TopazVideoEnhanceV2,
         ]
 
 

comfy_extras/nodes_textgen.py

@@ -32,6 +32,8 @@ class TextGenerate(io.ComfyNode):
                 io.Clip.Input("clip"),
                 io.String.Input("prompt", multiline=True, dynamic_prompts=True, default=""),
                 io.Image.Input("image", optional=True),
+                io.Image.Input("video", optional=True, tooltip="Video frames as image batch. Assumed to be 24 FPS; subsampled to 1 FPS internally."),
+                io.Audio.Input("audio", optional=True),
                 io.Int.Input("max_length", default=256, min=1, max=2048),
                 io.DynamicCombo.Input("sampling_mode", options=sampling_options, display_name="Sampling Mode"),
                 io.Boolean.Input("thinking", optional=True, default=False, tooltip="Operate in thinking mode if the model supports it."),
@@ -43,9 +45,9 @@ class TextGenerate(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True) -> io.NodeOutput:
+    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True, video=None, audio=None) -> io.NodeOutput:
 
-        tokens = clip.tokenize(prompt, image=image, skip_template=not use_default_template, min_length=1, thinking=thinking)
+        tokens = clip.tokenize(prompt, image=image, skip_template=not use_default_template, min_length=1, thinking=thinking, video=video, audio=audio)
 
         # Get sampling parameters from dynamic combo
         do_sample = sampling_mode.get("sampling_mode") == "on"
@@ -70,7 +72,8 @@ class TextGenerate(io.ComfyNode):
             seed=seed
         )
 
-        generated_text = clip.decode(generated_ids, skip_special_tokens=True)
+        generated_text = clip.decode(generated_ids)
+
         return io.NodeOutput(generated_text)
 
 
@@ -161,12 +164,12 @@ class TextGenerateLTX2Prompt(TextGenerate):
         )
 
     @classmethod
-    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True) -> io.NodeOutput:
+    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True, video=None, audio=None) -> io.NodeOutput:
         if image is None:
             formatted_prompt = f"<start_of_turn>system\n{LTX2_T2V_SYSTEM_PROMPT.strip()}<end_of_turn>\n<start_of_turn>user\nUser Raw Input Prompt: {prompt}.<end_of_turn>\n<start_of_turn>model\n"
         else:
             formatted_prompt = f"<start_of_turn>system\n{LTX2_I2V_SYSTEM_PROMPT.strip()}<end_of_turn>\n<start_of_turn>user\n\n<image_soft_token>\n\nUser Raw Input Prompt: {prompt}.<end_of_turn>\n<start_of_turn>model\n"
-        return super().execute(clip, formatted_prompt, max_length, sampling_mode, image, thinking, use_default_template)
+        return super().execute(clip, formatted_prompt, max_length, sampling_mode, image=image, thinking=thinking, use_default_template=use_default_template, video=video, audio=audio)
 
 
 class TextgenExtension(ComfyExtension):

execution.py

@@ -15,6 +15,7 @@ import torch
 from comfy.cli_args import args
 import comfy.memory_management
 import comfy.model_management
+import comfy.model_prefetch
 import comfy_aimdo.model_vbar
 
 from latent_preview import set_preview_method
@@ -537,6 +538,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                     if args.verbose == "DEBUG":
                         comfy_aimdo.control.analyze()
                     comfy.model_management.reset_cast_buffers()
+                    comfy.model_prefetch.cleanup_prefetch_queues()
                     comfy_aimdo.model_vbar.vbars_reset_watermark_limits()
 
             if has_pending_tasks:
@@ -1014,6 +1016,10 @@ async def validate_inputs(prompt_id, prompt, item, validated, visiting=None):
 
                 if isinstance(input_type, list) or input_type == io.Combo.io_type:
                     if input_type == io.Combo.io_type:
+                        # Skip validation for combos with remote options — options
+                        # are fetched client-side and not available on the server.
+                        if extra_info.get("remote_combo"):
+                            continue
                         combo_options = extra_info.get("options", [])
                     else:
                         combo_options = input_type

node_helpers.py

@@ -86,6 +86,6 @@ def image_alpha_fix(destination, source):
     if destination.shape[-1] < source.shape[-1]:
         source = source[...,:destination.shape[-1]]
     elif destination.shape[-1] > source.shape[-1]:
-        destination = torch.nn.functional.pad(destination, (0, 1))
-        destination[..., -1] = 1.0
+        source = torch.nn.functional.pad(source, (0, 1))
+        source[..., -1] = 1.0
     return destination, source

nodes.py

@@ -1694,26 +1694,27 @@ class LoadImage:
 
     RETURN_TYPES = ("IMAGE", "MASK")
     FUNCTION = "load_image"
+
     def load_image(self, image):
         image_path = folder_paths.get_annotated_filepath(image)
 
+        dtype = comfy.model_management.intermediate_dtype()
+        device = comfy.model_management.intermediate_device()
+
         components = InputImpl.VideoFromFile(image_path).get_components()
         if components.images.shape[0] > 0:
-            return (components.images, 1.0 - components.alpha[..., -1] if components.alpha is not None else torch.zeros((components.images.shape[0], 64, 64), dtype=torch.float32, device="cpu"))
+            return (components.images.to(device=device, dtype=dtype), (1.0 - components.alpha[..., -1]).to(device=device, dtype=dtype) if components.alpha is not None else torch.zeros((components.images.shape[0], 64, 64), dtype=dtype, device=device))
 
+        # This code is left here to handle animated webp which pyav does not support loading
         img = node_helpers.pillow(Image.open, image_path)
 
         output_images = []
         output_masks = []
         w, h = None, None
 
-        dtype = comfy.model_management.intermediate_dtype()
-
         for i in ImageSequence.Iterator(img):
             i = node_helpers.pillow(ImageOps.exif_transpose, i)
 
-            if i.mode == 'I':
-                i = i.point(lambda i: i * (1 / 255))
             image = i.convert("RGB")
 
             if len(output_images) == 0:
@@ -1728,25 +1729,15 @@ class LoadImage:
             if 'A' in i.getbands():
                 mask = np.array(i.getchannel('A')).astype(np.float32) / 255.0
                 mask = 1. - torch.from_numpy(mask)
-            elif i.mode == 'P' and 'transparency' in i.info:
-                mask = np.array(i.convert('RGBA').getchannel('A')).astype(np.float32) / 255.0
-                mask = 1. - torch.from_numpy(mask)
             else:
-                mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
+                mask = torch.zeros((64, 64), dtype=torch.float32, device="cpu")
             output_images.append(image.to(dtype=dtype))
             output_masks.append(mask.unsqueeze(0).to(dtype=dtype))
 
-            if img.format == "MPO":
-                break  # ignore all frames except the first one for MPO format
+        output_image = torch.cat(output_images, dim=0)
+        output_mask = torch.cat(output_masks, dim=0)
 
-        if len(output_images) > 1:
-            output_image = torch.cat(output_images, dim=0)
-            output_mask = torch.cat(output_masks, dim=0)
-        else:
-            output_image = output_images[0]
-            output_mask = output_masks[0]
-
-        return (output_image, output_mask)
+        return (output_image.to(device=device, dtype=dtype), output_mask.to(device=device, dtype=dtype))
 
     @classmethod
     def IS_CHANGED(s, image):

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.42.15
-comfyui-workflow-templates==0.9.66
+comfyui-workflow-templates==0.9.68
 comfyui-embedded-docs==0.4.4
 torch
 torchsde

tests-unit/comfy_api_test/remote_combo_options_test.py

@@ -0,0 +1,165 @@
+import pytest
+
+from comfy_api.latest._io import (
+    Combo,
+    RemoteComboOptions,
+    RemoteItemSchema,
+    RemoteOptions,
+)
+
+
+def _schema(**overrides):
+    defaults = dict(value_field="id", label_field="name")
+    return RemoteItemSchema(**{**defaults, **overrides})
+
+
+def _combo(**overrides):
+    defaults = dict(route="/proxy/foo", item_schema=_schema())
+    return RemoteComboOptions(**{**defaults, **overrides})
+
+
+def test_item_schema_defaults_accepted():
+    d = _schema().as_dict()
+    assert d == {"value_field": "id", "label_field": "name", "preview_type": "image"}
+
+
+def test_item_schema_full_config_accepted():
+    d = _schema(
+        preview_url_field="preview",
+        preview_type="audio",
+        description_field="desc",
+        search_fields=["first", "last", "profile.email"],
+    ).as_dict()
+    assert d["preview_type"] == "audio"
+    assert d["search_fields"] == ["first", "last", "profile.email"]
+
+
+@pytest.mark.parametrize(
+    "bad_fields",
+    [
+        ["{first} {last}"],
+        ["name", "{age}"],
+        ["leading{"],
+        ["trailing}"],
+    ],
+)
+def test_item_schema_rejects_template_strings_in_search_fields(bad_fields):
+    with pytest.raises(ValueError, match="search_fields"):
+        _schema(search_fields=bad_fields)
+
+
+@pytest.mark.parametrize("bad_preview_type", ["middle", "IMAGE", "", "gif"])
+def test_item_schema_rejects_unknown_preview_type(bad_preview_type):
+    with pytest.raises(ValueError, match="preview_type"):
+        _schema(preview_type=bad_preview_type)
+
+
+def test_combo_options_minimal_accepted():
+    d = _combo().as_dict()
+    assert d["route"] == "/proxy/foo"
+    assert d["refresh_button"] is True
+    assert "item_schema" in d
+
+
+@pytest.mark.parametrize(
+    "route",
+    [
+        "/proxy/foo",
+        "http://localhost:9000/voices",
+        "https://api.example.com/v1/voices",
+    ],
+)
+def test_combo_options_accepts_valid_routes(route):
+    _combo(route=route)
+
+
+@pytest.mark.parametrize(
+    "route",
+    [
+        "",
+        "api.example.com/voices",
+        "voices",
+        "ftp-no-scheme",
+    ],
+)
+def test_combo_options_rejects_invalid_routes(route):
+    with pytest.raises(ValueError, match="'route'"):
+        _combo(route=route)
+
+
+@pytest.mark.parametrize(
+    "route",
+    [
+        "http://localhost:9000/voices",
+        "https://api.example.com/v1/voices",
+    ],
+)
+def test_combo_options_use_comfy_api_rejects_absolute_route(route):
+    with pytest.raises(ValueError, match="use_comfy_api"):
+        _combo(route=route, use_comfy_api=True)
+
+
+def test_combo_options_use_comfy_api_accepts_relative_route():
+    _combo(route="/proxy/foo", use_comfy_api=True)
+
+
+def test_combo_options_page_size_with_response_key_rejected():
+    with pytest.raises(ValueError, match="response_key.*page_size"):
+        _combo(page_size=50, response_key="data.items")
+
+
+@pytest.mark.parametrize("bad_page_size", [0, -1])
+def test_combo_options_page_size_must_be_positive(bad_page_size):
+    with pytest.raises(ValueError, match="page_size"):
+        _combo(page_size=bad_page_size)
+
+
+@pytest.mark.parametrize("bad_auto_select", ["middle", "FIRST", "", "firstlast"])
+def test_combo_options_rejects_unknown_auto_select(bad_auto_select):
+    with pytest.raises(ValueError, match="auto_select"):
+        _combo(auto_select=bad_auto_select)
+
+
+@pytest.mark.parametrize("bad_refresh", [1, 127])
+def test_combo_options_refresh_in_forbidden_range_rejected(bad_refresh):
+    with pytest.raises(ValueError, match="refresh"):
+        _combo(refresh=bad_refresh)
+
+
+@pytest.mark.parametrize("ok_refresh", [0, -1, 128])
+def test_combo_options_refresh_valid_values_accepted(ok_refresh):
+    _combo(refresh=ok_refresh)
+
+
+def test_combo_options_timeout_negative_rejected():
+    with pytest.raises(ValueError, match="timeout"):
+        _combo(timeout=-1)
+
+
+def test_combo_options_max_retries_negative_rejected():
+    with pytest.raises(ValueError, match="max_retries"):
+        _combo(max_retries=-1)
+
+
+def test_combo_options_as_dict_prunes_none_fields():
+    d = _combo().as_dict()
+    for pruned in ("page_size", "response_key", "refresh", "timeout", "max_retries", "auto_select"):
+        assert pruned not in d
+
+
+def test_combo_input_accepts_remote_combo_alone():
+    Combo.Input("voice", remote_combo=_combo())
+
+
+def test_combo_input_rejects_remote_plus_remote_combo():
+    with pytest.raises(ValueError, match="remote.*remote_combo"):
+        Combo.Input(
+            "voice",
+            remote=RemoteOptions(route="/r", refresh_button=True),
+            remote_combo=_combo(),
+        )
+
+
+def test_combo_input_rejects_options_plus_remote_combo():
+    with pytest.raises(ValueError, match="options.*remote_combo"):
+        Combo.Input("voice", options=["a", "b"], remote_combo=_combo())