Merge d56887ac52 into 783782d5d7

* mm: Use Aimdo raw allocator for cast buffers

pytorch manages allocation of growing buffers on streams poorly. Pyt
has no windows support for the expandable segments allocator (which is
the right tool for this job), while also segmenting the memory by
stream such that it can be generally re-used. So kick the problem to
aimdo which can just grow a virtual region thats freed per stream.

* plan

* ops: move cpu handler up to the caller

* ops: split up prefetch from weight prep block prefetching API

Split up the casting and weight formating/lora stuff in prep for
arbitrary prefetch support.

* ops: implement block prefetching API

allow a model to construct a prefetch list and operate it for increased
async offload.

* ltxv2: Implement block prefetching

* Implement lora async offload

Implement async offload of loras.

.ci/windows_amd_base_files/run_amd_gpu_enable_dynamic_vram.bat

@@ -1,2 +1,2 @@
-.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-dynamic-vram
 pause

CODEOWNERS

@@ -1,2 +1,2 @@
 # Admins
-* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128
+* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128 @kijai

README.md

@@ -193,13 +193,15 @@ If you have trouble extracting it, right click the file -> properties -> unblock
 
 The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start.
 
-#### Alternative Downloads:
+#### All Official Portable Downloads:
 
 [Portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
 
-[Experimental portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
+[Portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
 
-[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
+[Portable for Nvidia GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z) (supports 20 series and above).
+
+[Portable for Nvidia GPUs with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
 
 #### How do I share models between another UI and ComfyUI?
 

comfy/cli_args.py

@@ -90,7 +90,6 @@ parser.add_argument("--force-channels-last", action="store_true", help="Force ch
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
-parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
 
 class LatentPreviewMethod(enum.Enum):

comfy/latent_formats.py

@@ -786,8 +786,26 @@ class ZImagePixelSpace(ChromaRadiance):
     pass
 
 class CogVideoX(LatentFormat):
+    """Latent format for CogVideoX-2b (THUDM/CogVideoX-2b).
+
+    scale_factor matches the vae/config.json scaling_factor for the 2b variant.
+    The 5b-class checkpoints (CogVideoX-5b, CogVideoX-1.5-5B, CogVideoX-Fun-V1.5-*)
+    use a different value; see CogVideoX1_5 below.
+    """
     latent_channels = 16
     latent_dimensions = 3
 
     def __init__(self):
         self.scale_factor = 1.15258426
+
+
+class CogVideoX1_5(CogVideoX):
+    """Latent format for 5b-class CogVideoX checkpoints.
+
+    Covers THUDM/CogVideoX-5b, THUDM/CogVideoX-1.5-5B, and the CogVideoX-Fun
+    V1.5-5b family (including VOID inpainting). All of these have
+    scaling_factor=0.7 in their vae/config.json. Auto-selected in
+    supported_models.CogVideoX_T2V based on transformer hidden dim.
+    """
+    def __init__(self):
+        self.scale_factor = 0.7

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/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
@@ -112,10 +113,6 @@ if args.directml is not None:
     # torch_directml.disable_tiled_resources(True)
     lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
 
-try:
-    import intel_extension_for_pytorch as ipex  # noqa: F401
-except:
-    pass
 
 try:
     _ = torch.xpu.device_count()
@@ -583,9 +580,6 @@ class LoadedModel:
 
         real_model = self.model.model
 
-        if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None:
-            with torch.no_grad():
-                real_model = ipex.optimize(real_model.eval(), inplace=True, graph_mode=True, concat_linear=True)
 
         self.real_model = weakref.ref(real_model)
         self.model_finalizer = weakref.finalize(real_model, cleanup_models)
@@ -1182,6 +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
@@ -1215,13 +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):
@@ -1581,10 +1592,7 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 3):
-            return True
-        else:
-            return torch.xpu.get_device_properties(device).has_fp16
+        return torch.xpu.get_device_properties(device).has_fp16
 
     if is_ascend_npu():
         return True
@@ -1650,10 +1658,7 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 3):
-            return True
-        else:
-            return torch.xpu.is_bf16_supported()
+        return torch.xpu.is_bf16_supported()
 
     if is_ascend_npu():
         return True
@@ -1784,6 +1789,7 @@ def soft_empty_cache(force=False):
     if cpu_state == CPUState.MPS:
         torch.mps.empty_cache()
     elif is_intel_xpu():
+        torch.xpu.synchronize()
         torch.xpu.empty_cache()
     elif is_ascend_npu():
         torch.npu.empty_cache()

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):

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.cogvideo
 
 import comfy.model_patcher
 import comfy.lora
@@ -1223,6 +1224,7 @@ class CLIPType(Enum):
     NEWBIE = 24
     FLUX2 = 25
     LONGCAT_IMAGE = 26
+    COGVIDEOX = 27
 
 
 
@@ -1418,6 +1420,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
                                                                         clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None)
                 clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
+            elif clip_type == CLIPType.COGVIDEOX:
+                clip_target.clip = comfy.text_encoders.cogvideo.cogvideo_te(**t5xxl_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.cogvideo.CogVideoXTokenizer
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer

comfy/supported_models.py

@@ -1853,6 +1853,14 @@ class CogVideoX_T2V(supported_models_base.BASE):
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
 
+    def __init__(self, unet_config):
+        # 2b-class (dim=1920, heads=30) uses scale_factor=1.15258426.
+        # 5b-class (dim=3072, heads=48) — incl. CogVideoX-5b, 1.5-5B, and
+        # Fun-V1.5 inpainting — uses scale_factor=0.7 per vae/config.json.
+        if unet_config.get("num_attention_heads", 0) >= 48:
+            self.latent_format = latent_formats.CogVideoX1_5
+        super().__init__(unet_config)
+
     def get_model(self, state_dict, prefix="", device=None):
         # CogVideoX 1.5 (patch_size_t=2) has different training base dimensions for RoPE
         if self.unet_config.get("patch_size_t") is not None:
@@ -1879,6 +1887,20 @@ class CogVideoX_I2V(CogVideoX_T2V):
         out = model_base.CogVideoX(self, image_to_video=True, device=device)
         return out
 
+class CogVideoX_Inpaint(CogVideoX_T2V):
+    unet_config = {
+        "image_model": "cogvideox",
+        "in_channels": 48,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, image_to_video=True, device=device)
+        return out
+
 
 models = [
     LotusD,
@@ -1958,6 +1980,7 @@ models = [
     ErnieImage,
     SAM3,
     SAM31,
+    CogVideoX_Inpaint,
     CogVideoX_I2V,
     CogVideoX_T2V,
     SVD_img2vid,

comfy/text_encoders/cogvideo.py

@@ -1,6 +1,48 @@
 import comfy.text_encoders.sd3_clip
+from comfy import sd1_clip
 
 
 class CogVideoXT5Tokenizer(comfy.text_encoders.sd3_clip.T5XXLTokenizer):
+    """Inner T5 tokenizer for CogVideoX.
+
+    CogVideoX was trained with T5 embeddings padded to 226 tokens (not 77 like SD3).
+    Used both directly by supported_models.CogVideoX_T2V.clip_target (paired with
+    the raw T5XXLModel) and by the CogVideoXTokenizer outer wrapper below.
+    """
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, min_length=226)
+
+
+class CogVideoXTokenizer(sd1_clip.SD1Tokenizer):
+    """Outer tokenizer wrapper for CLIPLoader (type="cogvideox")."""
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data,
+                         clip_name="t5xxl", tokenizer=CogVideoXT5Tokenizer)
+
+
+class CogVideoXT5XXL(sd1_clip.SD1ClipModel):
+    """Outer T5XXL model wrapper for CLIPLoader (type="cogvideox").
+
+    Wraps the raw T5XXL model in the SD1ClipModel interface so that CLIP.__init__
+    (which reads self.dtypes) works correctly. The inner model is the standard
+    sd3_clip.T5XXLModel (no attention_mask change needed for CogVideoX).
+    """
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="t5xxl",
+                         clip_model=comfy.text_encoders.sd3_clip.T5XXLModel,
+                         model_options=model_options)
+
+
+def cogvideo_te(dtype_t5=None, t5_quantization_metadata=None):
+    """Factory that returns a CogVideoXT5XXL class configured with the detected
+    T5 dtype and optional quantization metadata, for use in load_text_encoder_state_dicts.
+    """
+    class CogVideoXTEModel_(CogVideoXT5XXL):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
+            if dtype_t5 is not None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return CogVideoXTEModel_

comfy_extras/nodes_sdpose.py

@@ -459,27 +459,23 @@ class SDPoseKeypointExtractor(io.ComfyNode):
         total_images = image.shape[0]
         captured_feat = None
 
-        model_h = int(head.heatmap_size[0]) * 4   # e.g. 192 * 4 = 768
-        model_w = int(head.heatmap_size[1]) * 4   # e.g. 256 * 4 = 1024
+        model_w = int(head.heatmap_size[0]) * 4   # 192 * 4 = 768
+        model_h = int(head.heatmap_size[1]) * 4   # 256 * 4 = 1024
 
         def _resize_to_model(imgs):
-            """Aspect-preserving resize + zero-pad BHWC images to (model_h, model_w). Returns (resized_bhwc, scale, pad_top, pad_left)."""
+            """Stretch BHWC images to (model_h, model_w), model expects no aspect preservation."""
             h, w = imgs.shape[-3], imgs.shape[-2]
-            scale = min(model_h / h, model_w / w)
-            sh, sw = int(round(h * scale)), int(round(w * scale))
-            pt, pl = (model_h - sh) // 2, (model_w - sw) // 2
+            method = "area" if (model_h <= h and model_w <= w) else "bilinear"
             chw = imgs.permute(0, 3, 1, 2).float()
-            scaled = comfy.utils.common_upscale(chw, sw, sh, upscale_method="bilinear", crop="disabled")
-            padded = torch.zeros(scaled.shape[0], scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
-            padded[:, :, pt:pt + sh, pl:pl + sw] = scaled
-            return padded.permute(0, 2, 3, 1), scale, pt, pl
+            scaled = comfy.utils.common_upscale(chw, model_w, model_h, upscale_method=method, crop="disabled")
+            return scaled.permute(0, 2, 3, 1), model_w / w, model_h / h
 
-        def _remap_keypoints(kp, scale, pad_top, pad_left, offset_x=0, offset_y=0):
+        def _remap_keypoints(kp, scale_x, scale_y, offset_x=0, offset_y=0):
             """Remap keypoints from model space back to original image space."""
             kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
             invalid = kp[..., 0] < 0
-            kp[..., 0] = (kp[..., 0] - pad_left) / scale + offset_x
-            kp[..., 1] = (kp[..., 1] - pad_top)  / scale + offset_y
+            kp[..., 0] = kp[..., 0] / scale_x + offset_x
+            kp[..., 1] = kp[..., 1] / scale_y + offset_y
             kp[invalid] = -1
             return kp
 
@@ -529,18 +525,18 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                             continue
 
                         crop = img[:, y1:y2, x1:x2, :]  # (1, crop_h, crop_w, C)
-                        crop_resized, scale, pad_top, pad_left = _resize_to_model(crop)
+                        crop_resized, sx, sy = _resize_to_model(crop)
 
                         latent_crop = vae.encode(crop_resized)
                         kp_batch, sc_batch = _run_on_latent(latent_crop)
-                        kp = _remap_keypoints(kp_batch[0], scale, pad_top, pad_left, x1, y1)
+                        kp = _remap_keypoints(kp_batch[0], sx, sy, x1, y1)
                         img_keypoints.append(kp)
                         img_scores.append(sc_batch[0])
                 else:
-                    img_resized, scale, pad_top, pad_left = _resize_to_model(img)
+                    img_resized, sx, sy = _resize_to_model(img)
                     latent_img = vae.encode(img_resized)
                     kp_batch, sc_batch = _run_on_latent(latent_img)
-                    img_keypoints.append(_remap_keypoints(kp_batch[0], scale, pad_top, pad_left))
+                    img_keypoints.append(_remap_keypoints(kp_batch[0], sx, sy))
                     img_scores.append(sc_batch[0])
 
                 all_keypoints.append(img_keypoints)
@@ -549,12 +545,12 @@ class SDPoseKeypointExtractor(io.ComfyNode):
 
         else: # full-image mode, batched
             for batch_start in tqdm(range(0, total_images, batch_size), desc="Extracting keypoints"):
-                batch_resized, scale, pad_top, pad_left = _resize_to_model(image[batch_start:batch_start + batch_size])
+                batch_resized, sx, sy = _resize_to_model(image[batch_start:batch_start + batch_size])
                 latent_batch = vae.encode(batch_resized)
                 kp_batch, sc_batch = _run_on_latent(latent_batch)
 
                 for kp, sc in zip(kp_batch, sc_batch):
-                    all_keypoints.append([_remap_keypoints(kp, scale, pad_top, pad_left)])
+                    all_keypoints.append([_remap_keypoints(kp, sx, sy)])
                     all_scores.append([sc])
 
                 pbar.update(len(kp_batch))
@@ -727,13 +723,13 @@ class CropByBBoxes(io.ComfyNode):
                 scale = min(output_width / crop_w, output_height / crop_h)
                 scaled_w = int(round(crop_w * scale))
                 scaled_h = int(round(crop_h * scale))
-                scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="bilinear", crop="disabled")
+                scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="area", crop="disabled")
                 pad_left = (output_width  - scaled_w) // 2
                 pad_top  = (output_height - scaled_h) // 2
                 resized = torch.zeros(1, num_ch, output_height, output_width, dtype=image.dtype, device=image.device)
                 resized[:, :, pad_top:pad_top + scaled_h, pad_left:pad_left + scaled_w] = scaled
             else:  # "stretch"
-                resized = comfy.utils.common_upscale(crop_chw, output_width, output_height, upscale_method="bilinear", crop="disabled")
+                resized = comfy.utils.common_upscale(crop_chw, output_width, output_height, upscale_method="area", crop="disabled")
             crops.append(resized)
 
         if not crops:

comfy_extras/nodes_void.py

@@ -0,0 +1,483 @@
+import logging
+
+import torch
+
+import comfy
+import comfy.model_management
+import comfy.model_patcher
+import comfy.samplers
+import comfy.utils
+import folder_paths
+import node_helpers
+import nodes
+from comfy.utils import model_trange as trange
+from comfy_api.latest import ComfyExtension, io
+from torchvision.models.optical_flow import raft_large
+from typing_extensions import override
+
+
+from comfy_extras.void_noise_warp import RaftOpticalFlow, get_noise_from_video
+
+OpticalFlow = io.Custom("OPTICAL_FLOW")
+
+TEMPORAL_COMPRESSION = 4
+PATCH_SIZE_T = 2
+
+
+def _valid_void_length(length: int) -> int:
+    """Round ``length`` down to a value that produces an even latent_t.
+
+    VOID / CogVideoX-Fun-V1.5 uses patch_size_t=2, so the VAE-encoded latent
+    must have an even temporal dimension. If latent_t is odd, the transformer
+    pad_to_patch_size circular-wraps an extra latent frame onto the end; after
+    the post-transformer crop the last real latent frame has been influenced
+    by the wrapped phantom frame, producing visible jitter and "disappearing"
+    subjects near the end of the decoded video. Rounding down fixes this.
+    """
+    latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+    if latent_t % PATCH_SIZE_T == 0:
+        return length
+    # Round latent_t down to the nearest multiple of PATCH_SIZE_T, then invert
+    # the ((length - 1) // TEMPORAL_COMPRESSION) + 1 formula. Floor at 1 frame
+    # so we never return a non-positive length.
+    target_latent_t = max(PATCH_SIZE_T, (latent_t // PATCH_SIZE_T) * PATCH_SIZE_T)
+    return (target_latent_t - 1) * TEMPORAL_COMPRESSION + 1
+
+
+class OpticalFlowLoader(io.ComfyNode):
+    """Load an optical flow model from ``models/optical_flow/``.
+
+    Only torchvision's RAFT-large format is recognized today (the model used
+    by VOIDWarpedNoise).  The checkpoint must be placed under
+    ``models/optical_flow/`` — ComfyUI never downloads optical-flow weights
+    at runtime.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="OpticalFlowLoader",
+            display_name="Load Optical Flow Model",
+            category="loaders",
+            inputs=[
+                io.Combo.Input(
+                    "model_name",
+                    options=folder_paths.get_filename_list("optical_flow"),
+                    tooltip=(
+                        "Optical flow model to load.  Files must be placed in the "
+                        "'optical_flow' folder.  Today only torchvision's "
+                        "raft_large.pth is supported."
+                    ),
+                ),
+            ],
+            outputs=[
+                OpticalFlow.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_name) -> io.NodeOutput:
+
+        model_path = folder_paths.get_full_path_or_raise("optical_flow", model_name)
+        sd = comfy.utils.load_torch_file(model_path, safe_load=True)
+
+        has_raft_keys = (
+            any(k.startswith("feature_encoder.") for k in sd)
+            and any(k.startswith("context_encoder.") for k in sd)
+            and any(k.startswith("update_block.") for k in sd)
+        )
+        if not has_raft_keys:
+            raise ValueError(
+                "Unrecognized optical flow model format: expected a torchvision "
+                "RAFT-large state dict with 'feature_encoder.', 'context_encoder.' "
+                "and 'update_block.' prefixes."
+            )
+
+        model = raft_large(weights=None, progress=False)
+        model.load_state_dict(sd)
+        model.eval().to(torch.float32)
+
+        patcher = comfy.model_patcher.ModelPatcher(
+            model,
+            load_device=comfy.model_management.get_torch_device(),
+            offload_device=comfy.model_management.unet_offload_device(),
+        )
+        return io.NodeOutput(patcher)
+
+
+class VOIDQuadmaskPreprocess(io.ComfyNode):
+    """Preprocess a quadmask video for VOID inpainting.
+
+    Quantizes mask values to four semantic levels, inverts, and normalizes:
+      0   -> primary object to remove
+      63  -> overlap of primary + affected
+      127 -> affected region (interactions)
+      255 -> background (keep)
+
+    After inversion and normalization, the output mask has values in [0, 1]
+    with four discrete levels: 1.0 (remove), ~0.75, ~0.50, 0.0 (keep).
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDQuadmaskPreprocess",
+            category="mask/video",
+            inputs=[
+                io.Mask.Input("mask"),
+                io.Int.Input("dilate_width", default=0, min=0, max=50, step=1,
+                             tooltip="Dilation radius for the primary mask region (0 = no dilation)"),
+            ],
+            outputs=[
+                io.Mask.Output(display_name="quadmask"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, mask, dilate_width=0) -> io.NodeOutput:
+        m = mask.clone()
+
+        if m.max() <= 1.0:
+            m = m * 255.0
+
+        if dilate_width > 0 and m.ndim >= 3:
+            binary = (m < 128).float()
+            kernel_size = dilate_width * 2 + 1
+            if binary.ndim == 3:
+                binary = binary.unsqueeze(1)
+            dilated = torch.nn.functional.max_pool2d(
+                binary, kernel_size=kernel_size, stride=1, padding=dilate_width
+            )
+            if dilated.ndim == 4:
+                dilated = dilated.squeeze(1)
+            m = torch.where(dilated > 0.5, torch.zeros_like(m), m)
+
+        m = torch.where(m <= 31, torch.zeros_like(m), m)
+        m = torch.where((m > 31) & (m <= 95), torch.full_like(m, 63), m)
+        m = torch.where((m > 95) & (m <= 191), torch.full_like(m, 127), m)
+        m = torch.where(m > 191, torch.full_like(m, 255), m)
+
+        m = (255.0 - m) / 255.0
+
+        return io.NodeOutput(m)
+
+
+class VOIDInpaintConditioning(io.ComfyNode):
+    """Build VOID inpainting conditioning for CogVideoX.
+
+    Encodes the processed quadmask and masked source video through the VAE,
+    producing a 32-channel concat conditioning (16ch mask + 16ch masked video)
+    that gets concatenated with the 16ch noise latent by the model.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDInpaintConditioning",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Image.Input("video", tooltip="Source video frames [T, H, W, 3]"),
+                io.Mask.Input("quadmask", tooltip="Preprocessed quadmask from VOIDQuadmaskPreprocess [T, H, W]"),
+                io.Int.Input("width", default=672, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("height", default=384, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("length", default=45, min=1, max=nodes.MAX_RESOLUTION, step=1,
+                             tooltip="Number of pixel frames to process. For CogVideoX-Fun-V1.5 "
+                                     "(patch_size_t=2), latent_t must be even — lengths that "
+                                     "produce odd latent_t are rounded down (e.g. 49 → 45)."),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, video, quadmask,
+                width, height, length, batch_size) -> io.NodeOutput:
+
+        adjusted_length = _valid_void_length(length)
+        if adjusted_length != length:
+            logging.warning(
+                "VOIDInpaintConditioning: rounding length %d down to %d so that "
+                "latent_t is even (required by CogVideoX-Fun-V1.5 patch_size_t=2). "
+                "Using odd latent_t causes the last frame to be corrupted by "
+                "circular padding.", length, adjusted_length,
+            )
+            length = adjusted_length
+
+        latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+        latent_h = height // 8
+        latent_w = width // 8
+
+        vid = video[:length]
+        vid = comfy.utils.common_upscale(
+            vid.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+
+        qm = quadmask[:length]
+        if qm.ndim == 3:
+            qm = qm.unsqueeze(-1)
+        qm = comfy.utils.common_upscale(
+            qm.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+        if qm.ndim == 4 and qm.shape[-1] == 1:
+            qm = qm.squeeze(-1)
+
+        mask_condition = qm
+        if mask_condition.ndim == 3:
+            mask_condition_3ch = mask_condition.unsqueeze(-1).expand(-1, -1, -1, 3)
+        else:
+            mask_condition_3ch = mask_condition
+
+        inverted_mask_3ch = 1.0 - mask_condition_3ch
+        masked_video = vid[:, :, :, :3] * (1.0 - mask_condition_3ch)
+
+        mask_latents = vae.encode(inverted_mask_3ch)
+        masked_video_latents = vae.encode(masked_video)
+
+        def _match_temporal(lat, target_t):
+            if lat.shape[2] > target_t:
+                return lat[:, :, :target_t]
+            elif lat.shape[2] < target_t:
+                pad = target_t - lat.shape[2]
+                return torch.cat([lat, lat[:, :, -1:].repeat(1, 1, pad, 1, 1)], dim=2)
+            return lat
+
+        mask_latents = _match_temporal(mask_latents, latent_t)
+        masked_video_latents = _match_temporal(masked_video_latents, latent_t)
+
+        inpaint_latents = torch.cat([mask_latents, masked_video_latents], dim=1)
+
+        # No explicit scaling needed here: the model's CogVideoX.concat_cond()
+        # applies process_latent_in (×latent_format.scale_factor) to each 16-ch
+        # block of the stored conditioning. For 5b-class checkpoints (incl. the
+        # VOID/CogVideoX-Fun-V1.5 inpainting model) that scale_factor is auto-
+        # selected as 0.7 in supported_models.CogVideoX_T2V, which matches the
+        # diffusers vae/config.json scaling_factor VOID was trained with.
+
+        positive = node_helpers.conditioning_set_values(
+            positive, {"concat_latent_image": inpaint_latents}
+        )
+        negative = node_helpers.conditioning_set_values(
+            negative, {"concat_latent_image": inpaint_latents}
+        )
+
+        noise_latent = torch.zeros(
+            [batch_size, 16, latent_t, latent_h, latent_w],
+            device=comfy.model_management.intermediate_device()
+        )
+
+        return io.NodeOutput(positive, negative, {"samples": noise_latent})
+
+
+class VOIDWarpedNoise(io.ComfyNode):
+    """Generate optical-flow warped noise for VOID Pass 2 refinement.
+
+    Takes the Pass 1 output video and produces temporally-correlated noise
+    by warping Gaussian noise along optical flow vectors. This noise is used
+    as the initial latent for Pass 2, resulting in better temporal consistency.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDWarpedNoise",
+            category="latent/video",
+            inputs=[
+                OpticalFlow.Input(
+                    "optical_flow",
+                    tooltip="Optical flow model from OpticalFlowLoader (RAFT-large).",
+                ),
+                io.Image.Input("video", tooltip="Pass 1 output video frames [T, H, W, 3]"),
+                io.Int.Input("width", default=672, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("height", default=384, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("length", default=45, min=1, max=nodes.MAX_RESOLUTION, step=1,
+                             tooltip="Number of pixel frames. Rounded down to make latent_t "
+                                     "even (patch_size_t=2 requirement), e.g. 49 → 45."),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Latent.Output(display_name="warped_noise"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, optical_flow, video, width, height, length, batch_size) -> io.NodeOutput:
+
+        adjusted_length = _valid_void_length(length)
+        if adjusted_length != length:
+            logging.warning(
+                "VOIDWarpedNoise: rounding length %d down to %d so that "
+                "latent_t is even (required by CogVideoX-Fun-V1.5 patch_size_t=2).",
+                length, adjusted_length,
+            )
+            length = adjusted_length
+
+        latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+        latent_h = height // 8
+        latent_w = width // 8
+
+        # RAFT + noise warp is real compute, not an "intermediate" buffer, so
+        # we want the actual torch device (CUDA/MPS).  The final latent is
+        # moved back to intermediate_device() before returning to match the
+        # rest of the ComfyUI pipeline.
+        device = comfy.model_management.get_torch_device()
+
+        comfy.model_management.load_model_gpu(optical_flow)
+        raft = RaftOpticalFlow(optical_flow.model, device=device)
+
+        vid = video[:length].to(device)
+        vid = comfy.utils.common_upscale(
+            vid.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+        vid_uint8 = (vid.clamp(0, 1) * 255).to(torch.uint8)
+
+        FRAME = 2**-1
+        FLOW = 2**3
+        LATENT_SCALE = 8
+
+        warped = get_noise_from_video(
+            vid_uint8,
+            raft,
+            noise_channels=16,
+            resize_frames=FRAME,
+            resize_flow=FLOW,
+            downscale_factor=round(FRAME * FLOW) * LATENT_SCALE,
+            device=device,
+        )
+
+        if warped.shape[0] != latent_t:
+            indices = torch.linspace(0, warped.shape[0] - 1, latent_t,
+                                     device=device).long()
+            warped = warped[indices]
+
+        if warped.shape[1] != latent_h or warped.shape[2] != latent_w:
+            # (T, H, W, C) → (T, C, H, W) → bilinear resize → back
+            warped = warped.permute(0, 3, 1, 2)
+            warped = torch.nn.functional.interpolate(
+                warped, size=(latent_h, latent_w),
+                mode="bilinear", align_corners=False,
+            )
+            warped = warped.permute(0, 2, 3, 1)
+
+        # (T, H, W, C) → (B, C, T, H, W)
+        warped_tensor = warped.permute(3, 0, 1, 2).unsqueeze(0)
+        if batch_size > 1:
+            warped_tensor = warped_tensor.repeat(batch_size, 1, 1, 1, 1)
+
+        warped_tensor = warped_tensor.to(comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": warped_tensor})
+
+
+class Noise_FromLatent:
+    """Wraps a pre-computed LATENT tensor as a NOISE source."""
+    def __init__(self, latent_dict):
+        self.seed = 0
+        self._samples = latent_dict["samples"]
+
+    def generate_noise(self, input_latent):
+        return self._samples.clone().cpu()
+
+
+class VOIDWarpedNoiseSource(io.ComfyNode):
+    """Convert a LATENT (e.g. from VOIDWarpedNoise) into a NOISE source
+    for use with SamplerCustomAdvanced."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDWarpedNoiseSource",
+            category="sampling/custom_sampling/noise",
+            inputs=[
+                io.Latent.Input("warped_noise",
+                    tooltip="Warped noise latent from VOIDWarpedNoise"),
+            ],
+            outputs=[io.Noise.Output()],
+        )
+
+    @classmethod
+    def execute(cls, warped_noise) -> io.NodeOutput:
+        return io.NodeOutput(Noise_FromLatent(warped_noise))
+
+
+class VOID_DDIM(comfy.samplers.Sampler):
+    """DDIM sampler for VOID inpainting models.
+
+    VOID was trained with the diffusers CogVideoXDDIMScheduler which operates in
+    alpha-space (input std ≈ 1). The standard KSampler applies noise_scaling that
+    multiplies by sqrt(1+sigma^2) ≈ 4500x, which is incompatible with VOID's
+    training. This sampler skips noise_scaling and implements the DDIM update rule
+    directly using sigma-to-alpha conversion.
+    """
+
+    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
+        x = noise.to(torch.float32)
+        model_options = extra_args.get("model_options", {})
+        seed = extra_args.get("seed", None)
+        s_in = x.new_ones([x.shape[0]])
+
+        for i in trange(len(sigmas) - 1, disable=disable_pbar):
+            sigma = sigmas[i]
+            sigma_next = sigmas[i + 1]
+
+            denoised = model_wrap(x, sigma * s_in, model_options=model_options, seed=seed)
+
+            if callback is not None:
+                callback(i, denoised, x, len(sigmas) - 1)
+
+            if sigma_next == 0:
+                x = denoised
+            else:
+                alpha_t = 1.0 / (1.0 + sigma ** 2)
+                alpha_prev = 1.0 / (1.0 + sigma_next ** 2)
+
+                pred_eps = (x - (alpha_t ** 0.5) * denoised) / (1.0 - alpha_t) ** 0.5
+                x = (alpha_prev ** 0.5) * denoised + (1.0 - alpha_prev) ** 0.5 * pred_eps
+
+        return x
+
+
+class VOIDSampler(io.ComfyNode):
+    """VOID DDIM sampler for use with SamplerCustom / SamplerCustomAdvanced.
+
+    Required for VOID inpainting models. Implements the same DDIM loop that VOID
+    was trained with (diffusers CogVideoXDDIMScheduler), without the noise_scaling
+    that the standard KSampler applies. Use with RandomNoise or VOIDWarpedNoiseSource.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDSampler",
+            category="sampling/custom_sampling/samplers",
+            inputs=[],
+            outputs=[io.Sampler.Output()],
+        )
+
+    @classmethod
+    def execute(cls) -> io.NodeOutput:
+        return io.NodeOutput(VOID_DDIM())
+
+    get_sampler = execute
+
+
+class VOIDExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            OpticalFlowLoader,
+            VOIDQuadmaskPreprocess,
+            VOIDInpaintConditioning,
+            VOIDWarpedNoise,
+            VOIDWarpedNoiseSource,
+            VOIDSampler,
+        ]
+
+
+async def comfy_entrypoint() -> VOIDExtension:
+    return VOIDExtension()

comfy_extras/void_noise_warp.py

@@ -0,0 +1,494 @@
+"""
+Optical-flow-warped noise for VOID Pass 2 refinement.
+
+Adapted from RyannDaGreat/CommonSource (MIT License, Ryan Burgert):
+  https://github.com/RyannDaGreat/CommonSource
+  - noise_warp.py  (NoiseWarper / warp_xyωc / regaussianize / get_noise_from_video)
+  - raft.py        (RaftOpticalFlow)
+
+Only the code paths that ``comfy_extras/nodes_void.py::VOIDWarpedNoise`` actually
+uses (torch THWC uint8 input, no background removal, no visualization, no disk
+I/O, default warp/noise params) have been inlined.  External ``rp`` utilities
+have been replaced with equivalents from torch.nn.functional / einops.  The
+RAFT optical-flow model itself is loaded offline via ``OpticalFlowLoader`` in
+``nodes_void.py`` and passed into ``get_noise_from_video`` by the caller; this
+module never downloads weights at runtime.
+"""
+
+import logging
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+
+import comfy.model_management
+
+
+# ---------------------------------------------------------------------------
+# Low-level torch image helpers (drop-in replacements for rp.torch_* primitives)
+# ---------------------------------------------------------------------------
+
+def _torch_resize_chw(image, size, interp, copy=True):
+    """Resize a CHW tensor.
+
+    ``size`` is either a scalar factor or a (h, w) tuple.  ``interp`` is one
+    of ``"bilinear"``, ``"nearest"``, ``"area"``.  When ``copy`` is False and
+    the requested size matches the input, returns the input tensor as is
+    (faster but callers must not mutate the result).
+    """
+    if image.ndim != 3:
+        raise ValueError(
+            f"_torch_resize_chw expects a 3D CHW tensor, got shape {tuple(image.shape)}"
+        )
+    _, in_h, in_w = image.shape
+    if isinstance(size, (int, float)) and not isinstance(size, bool):
+        new_h = max(1, int(in_h * size))
+        new_w = max(1, int(in_w * size))
+    else:
+        new_h, new_w = size
+
+    if (new_h, new_w) == (in_h, in_w):
+        return image.clone() if copy else image
+
+    kwargs = {}
+    if interp in ("bilinear", "bicubic"):
+        kwargs["align_corners"] = False
+    out = F.interpolate(image[None], size=(new_h, new_w), mode=interp, **kwargs)[0]
+    return out
+
+
+def _torch_remap_relative(image, dx, dy, interp="bilinear"):
+    """Relative remap of a CHW image via ``F.grid_sample``.
+
+    Equivalent to ``rp.torch_remap_image(image, dx, dy, relative=True, interp=interp)``
+    for ``interp`` in {"bilinear", "nearest"}.  Out-of-bounds samples are 0.
+    """
+    if image.ndim != 3:
+        raise ValueError(
+            f"_torch_remap_relative expects a 3D CHW tensor, got shape {tuple(image.shape)}"
+        )
+    if dx.shape != dy.shape:
+        raise ValueError(
+            f"_torch_remap_relative: dx and dy must match, got {tuple(dx.shape)} vs {tuple(dy.shape)}"
+        )
+    _, h, w = image.shape
+
+    x_abs = dx + torch.arange(w, device=dx.device, dtype=dx.dtype)
+    y_abs = dy + torch.arange(h, device=dy.device, dtype=dy.dtype)[:, None]
+
+    x_norm = (x_abs / (w - 1)) * 2 - 1
+    y_norm = (y_abs / (h - 1)) * 2 - 1
+
+    grid = torch.stack([x_norm, y_norm], dim=-1)[None].to(image.dtype)
+    out = F.grid_sample(
+        image[None], grid, mode=interp, align_corners=True, padding_mode="zeros"
+    )[0]
+    return out
+
+
+def _torch_scatter_add_relative(image, dx, dy):
+    """Scatter-add a CHW image using relative floor-rounded (dx, dy) offsets.
+
+    Equivalent to ``rp.torch_scatter_add_image(image, dx, dy, relative=True,
+    interp='floor')``.  Out-of-bounds targets are dropped.
+    """
+    if image.ndim != 3:
+        raise ValueError(
+            f"_torch_scatter_add_relative expects a 3D CHW tensor, got shape {tuple(image.shape)}"
+        )
+    in_c, in_h, in_w = image.shape
+    if dx.shape != (in_h, in_w) or dy.shape != (in_h, in_w):
+        raise ValueError(
+            f"_torch_scatter_add_relative: dx/dy must be ({in_h}, {in_w}), "
+            f"got dx={tuple(dx.shape)} dy={tuple(dy.shape)}"
+        )
+
+    x = dx.long() + torch.arange(in_w, device=dx.device, dtype=torch.long)
+    y = dy.long() + torch.arange(in_h, device=dy.device, dtype=torch.long)[:, None]
+
+    valid = ((y >= 0) & (y < in_h) & (x >= 0) & (x < in_w)).reshape(-1)
+    indices = (y * in_w + x).reshape(-1)[valid]
+
+    flat_image = rearrange(image, "c h w -> (h w) c")[valid]
+    out = torch.zeros((in_h * in_w, in_c), dtype=image.dtype, device=image.device)
+    out.index_add_(0, indices, flat_image)
+    return rearrange(out, "(h w) c -> c h w", h=in_h, w=in_w)
+
+
+# ---------------------------------------------------------------------------
+# Noise warping primitives (ported from noise_warp.py)
+# ---------------------------------------------------------------------------
+
+def unique_pixels(image):
+    """Find unique pixel values in a CHW tensor.
+
+    Returns ``(unique_colors [U, C], counts [U], index_matrix [H, W])`` where
+    ``index_matrix[i, j]`` is the index of the unique color at that pixel.
+    """
+    _, h, w = image.shape
+    flat = rearrange(image, "c h w -> (h w) c")
+    unique_colors, inverse_indices, counts = torch.unique(
+        flat, dim=0, return_inverse=True, return_counts=True, sorted=False,
+    )
+    index_matrix = rearrange(inverse_indices, "(h w) -> h w", h=h, w=w)
+    return unique_colors, counts, index_matrix
+
+
+def sum_indexed_values(image, index_matrix):
+    """For each unique index, sum the CHW image values at its pixels."""
+    _, h, w = image.shape
+    u = int(index_matrix.max().item()) + 1
+    flat = rearrange(image, "c h w -> (h w) c")
+    out = torch.zeros((u, flat.shape[1]), dtype=flat.dtype, device=flat.device)
+    out.index_add_(0, index_matrix.view(-1), flat)
+    return out
+
+
+def indexed_to_image(index_matrix, unique_colors):
+    """Build a CHW image from an index matrix and a (U, C) color table."""
+    h, w = index_matrix.shape
+    flat = unique_colors[index_matrix.view(-1)]
+    return rearrange(flat, "(h w) c -> c h w", h=h, w=w)
+
+
+def regaussianize(noise):
+    """Variance-preserving re-sampling of a CHW noise tensor.
+
+    Wherever the noise contains groups of identical pixel values (e.g. after
+    a nearest-neighbor warp that duplicated source pixels), adds zero-mean
+    foreign noise within each group and scales by ``1/sqrt(count)`` so the
+    output is unit-variance gaussian again.
+    """
+    _, hs, ws = noise.shape
+    _, counts, index_matrix = unique_pixels(noise[:1])
+
+    foreign_noise = torch.randn_like(noise)
+    summed = sum_indexed_values(foreign_noise, index_matrix)
+    meaned = indexed_to_image(index_matrix, summed / rearrange(counts, "u -> u 1"))
+    zeroed_foreign = foreign_noise - meaned
+
+    counts_image = indexed_to_image(index_matrix, rearrange(counts, "u -> u 1"))
+
+    output = noise / counts_image ** 0.5 + zeroed_foreign
+    return output, counts_image
+
+
+def xy_meshgrid_like_image(image):
+    """Return a (2, H, W) tensor of (x, y) pixel coordinates matching ``image``."""
+    _, h, w = image.shape
+    y, x = torch.meshgrid(
+        torch.arange(h, device=image.device, dtype=image.dtype),
+        torch.arange(w, device=image.device, dtype=image.dtype),
+        indexing="ij",
+    )
+    return torch.stack([x, y])
+
+
+def noise_to_state(noise):
+    """Pack a (C, H, W) noise tensor into a state tensor (3+C, H, W) = [dx, dy, ω, noise]."""
+    zeros = torch.zeros_like(noise[:1])
+    ones = torch.ones_like(noise[:1])
+    return torch.cat([zeros, zeros, ones, noise])
+
+
+def state_to_noise(state):
+    """Unpack the noise channels from a state tensor."""
+    return state[3:]
+
+
+def warp_state(state, flow):
+    """Warp a noise-warper state tensor along the given optical flow.
+
+    ``state`` has shape ``(3+c, h, w)`` (= dx, dy, ω, c noise channels).
+    ``flow`` has shape ``(2, h, w)`` (= dx, dy).
+    """
+    if flow.device != state.device:
+        raise ValueError(
+            f"warp_state: flow and state must be on the same device, "
+            f"got flow={flow.device} state={state.device}"
+        )
+    if state.ndim != 3:
+        raise ValueError(
+            f"warp_state: state must be 3D (3+C, H, W), got shape {tuple(state.shape)}"
+        )
+    xyoc, h, w = state.shape
+    if flow.shape != (2, h, w):
+        raise ValueError(
+            f"warp_state: flow must have shape (2, {h}, {w}), got {tuple(flow.shape)}"
+        )
+    device = state.device
+
+    x_ch, y_ch = 0, 1
+    xy = 2         # state[:xy]  = [dx, dy]
+    xyw = 3        # state[:xyw] = [dx, dy, ω]
+    w_ch = 2       # state[w_ch] = ω
+    c = xyoc - xyw
+    oc = xyoc - xy
+    if c <= 0:
+        raise ValueError(
+            f"warp_state: state has no noise channels (expected 3+C with C>0, got {xyoc} channels)"
+        )
+    if not (state[w_ch] > 0).all():
+        raise ValueError("warp_state: all weights in state[2] must be > 0")
+
+    grid = xy_meshgrid_like_image(state)
+
+    init = torch.empty_like(state)
+    init[:xy] = 0
+    init[w_ch] = 1
+    init[-c:] = 0
+
+    # --- Expansion branch: nearest-neighbor remap with negated flow ---
+    pre_expand = torch.empty_like(state)
+    pre_expand[:xy] = _torch_remap_relative(state[:xy], -flow[0], -flow[1], "nearest")
+    pre_expand[-oc:] = _torch_remap_relative(state[-oc:], -flow[0], -flow[1], "nearest")
+    pre_expand[w_ch][pre_expand[w_ch] == 0] = 1
+
+    # --- Shrink branch: scatter-add state into new positions ---
+    pre_shrink = state.clone()
+    pre_shrink[:xy] += flow
+
+    pos = (grid + pre_shrink[:xy]).round()
+    in_bounds = (pos[x_ch] >= 0) & (pos[x_ch] < w) & (pos[y_ch] >= 0) & (pos[y_ch] < h)
+    pre_shrink = torch.where(~in_bounds[None], init, pre_shrink)
+
+    scat_xy = pre_shrink[:xy].round()
+    pre_shrink[:xy] -= scat_xy
+    pre_shrink[:xy] = 0  # xy_mode='none' in upstream
+
+    def scat(tensor):
+        return _torch_scatter_add_relative(tensor, scat_xy[0], scat_xy[1])
+
+    # rp.torch_scatter_add_image on a bool tensor errors on modern torch;
+    # scatter-sum a float ones tensor and threshold to get the mask instead.
+    shrink_mask = scat(torch.ones(1, h, w, dtype=state.dtype, device=device)) > 0
+
+    # Drop expansion samples at positions that will be filled by shrink.
+    pre_expand = torch.where(shrink_mask, init, pre_expand)
+
+    # Regaussianize both branches together so duplicated-source groups are
+    # counted globally, then split back apart.
+    concat = torch.cat([pre_shrink, pre_expand], dim=2)  # along width
+    concat[-c:], counts_image = regaussianize(concat[-c:])
+    concat[w_ch] = concat[w_ch] / counts_image[0]
+    concat[w_ch] = concat[w_ch].nan_to_num()
+    pre_shrink, expand = torch.chunk(concat, chunks=2, dim=2)
+
+    shrink = torch.empty_like(pre_shrink)
+    shrink[w_ch] = scat(pre_shrink[w_ch][None])[0]
+    shrink[:xy] = scat(pre_shrink[:xy] * pre_shrink[w_ch][None]) / shrink[w_ch][None]
+    shrink[-c:] = scat(pre_shrink[-c:] * pre_shrink[w_ch][None]) / scat(
+        pre_shrink[w_ch][None] ** 2
+    ).sqrt()
+
+    output = torch.where(shrink_mask, shrink, expand)
+    output[w_ch] = output[w_ch] / output[w_ch].mean()
+    output[w_ch] += 1e-5
+    output[w_ch] **= 0.9999
+    return output
+
+
+class NoiseWarper:
+    """Maintain a warpable noise state and emit gaussian noise per frame.
+
+    Simplified from RyannDaGreat/CommonSource/noise_warp.py::NoiseWarper:
+    ``scale_factor``, ``post_noise_alpha``, ``progressive_noise_alpha``, and
+    ``warp_kwargs`` are all dropped since VOIDWarpedNoise always uses defaults.
+    """
+
+    def __init__(self, c, h, w, device, dtype=torch.float32):
+        if c <= 0 or h <= 0 or w <= 0:
+            raise ValueError(
+                f"NoiseWarper: c/h/w must all be positive, got c={c} h={h} w={w}"
+            )
+        self.c = c
+        self.h = h
+        self.w = w
+        self.device = device
+        self.dtype = dtype
+
+        noise = torch.randn(c, h, w, dtype=dtype, device=device)
+        self._state = noise_to_state(noise)
+
+    @property
+    def noise(self):
+        # With scale_factor=1 the "downsample to respect weights" step is a
+        # size-preserving no-op; the weight-variance correction math still
+        # runs to stay faithful to upstream.
+        n = state_to_noise(self._state)
+        weights = self._state[2:3]
+        return n * weights / (weights ** 2).sqrt()
+
+    def __call__(self, dx, dy):
+        if dx.shape != dy.shape:
+            raise ValueError(
+                f"NoiseWarper: dx and dy must match, got {tuple(dx.shape)} vs {tuple(dy.shape)}"
+            )
+        flow = torch.stack([dx, dy]).to(self.device, self.dtype)
+        _, oflowh, ofloww = flow.shape
+
+        flow = _torch_resize_chw(flow, (self.h, self.w), "bilinear", copy=True)
+        flowh, floww = flow.shape[-2:]
+
+        # Upstream scales flow[0] by flowh/oflowh and flow[1] by floww/ofloww
+        # (channel-order appears swapped but harmless when H and W are scaled
+        # by the same factor, which is always the case for our callers).
+        flow[0] *= flowh / oflowh
+        flow[1] *= floww / ofloww
+
+        self._state = warp_state(self._state, flow)
+        return self
+
+
+# ---------------------------------------------------------------------------
+# RAFT optical flow wrapper (ported from raft.py)
+# ---------------------------------------------------------------------------
+
+class RaftOpticalFlow:
+    """RAFT-large wrapper around a pre-loaded torchvision model.
+
+    ``model`` must be the ``torchvision.models.optical_flow.raft_large`` module
+    with its weights already populated; this class is load-agnostic so the
+    caller owns downloading/offload concerns (see ``OpticalFlowLoader`` in
+    ``nodes_void.py``).  ``__call__`` returns a ``(2, H, W)`` flow.
+    """
+
+    def __init__(self, model, device=None):
+        if device is None:
+            device = comfy.model_management.get_torch_device()
+        device = torch.device(device) if not isinstance(device, torch.device) else device
+
+        model = model.to(device)
+        model.eval()
+        self.device = device
+        self.model = model
+
+    def _preprocess(self, image_chw):
+        image = image_chw.to(self.device, torch.float32)
+        _, h, w = image.shape
+        new_h = (h // 8) * 8
+        new_w = (w // 8) * 8
+        image = _torch_resize_chw(image, (new_h, new_w), "bilinear", copy=False)
+        image = image * 2 - 1
+        return image[None]
+
+    def __call__(self, from_image, to_image):
+        """``from_image``, ``to_image``: CHW float tensors in [0, 1]."""
+        if from_image.shape != to_image.shape:
+            raise ValueError(
+                f"RaftOpticalFlow: from_image and to_image must match, "
+                f"got {tuple(from_image.shape)} vs {tuple(to_image.shape)}"
+            )
+        _, h, w = from_image.shape
+        with torch.no_grad():
+            img1 = self._preprocess(from_image)
+            img2 = self._preprocess(to_image)
+            list_of_flows = self.model(img1, img2)
+            flow = list_of_flows[-1][0]  # (2, new_h, new_w)
+            if flow.shape[-2:] != (h, w):
+                flow = _torch_resize_chw(flow, (h, w), "bilinear", copy=False)
+        return flow
+
+
+# ---------------------------------------------------------------------------
+# Narrow entry point used by VOIDWarpedNoise
+# ---------------------------------------------------------------------------
+
+def get_noise_from_video(
+    video_frames: torch.Tensor,
+    raft: RaftOpticalFlow,
+    *,
+    noise_channels: int = 16,
+    resize_frames: float = 0.5,
+    resize_flow: int = 8,
+    downscale_factor: int = 32,
+    device: Optional[torch.device] = None,
+) -> torch.Tensor:
+    """Produce optical-flow-warped gaussian noise from a video.
+
+    Args:
+        video_frames: ``(T, H, W, 3)`` uint8 torch tensor.
+        raft: Pre-loaded RAFT optical-flow wrapper (see ``RaftOpticalFlow``).
+        noise_channels: Channels in the output noise.
+        resize_frames: Pre-RAFT frame scale factor.
+        resize_flow: Post-flow up-scale factor applied to the optical flow;
+            the internal noise state is allocated at
+            ``(resize_flow * resize_frames * H, resize_flow * resize_frames * W)``.
+        downscale_factor: Area-pool factor applied to the noise before return;
+            should evenly divide the internal noise resolution.
+        device: Target device.  Defaults to ``comfy.model_management.get_torch_device()``.
+
+    Returns:
+        ``(T, H', W', noise_channels)`` float32 noise tensor on ``device``.
+    """
+    if not isinstance(resize_flow, int) or resize_flow < 1:
+        raise ValueError(
+            f"get_noise_from_video: resize_flow must be a positive int, got {resize_flow!r}"
+        )
+    if video_frames.ndim != 4 or video_frames.shape[-1] != 3:
+        raise ValueError(
+            "get_noise_from_video: video_frames must have shape (T, H, W, 3), "
+            f"got {tuple(video_frames.shape)}"
+        )
+    if video_frames.dtype != torch.uint8:
+        raise TypeError(
+            "get_noise_from_video: video_frames must be uint8 in [0, 255], "
+            f"got dtype {video_frames.dtype}"
+        )
+
+    if device is None:
+        device = comfy.model_management.get_torch_device()
+    device = torch.device(device) if not isinstance(device, torch.device) else device
+
+    if device.type == "cpu":
+        logging.warning(
+            "VOIDWarpedNoise: running get_noise_from_video on CPU; this will be "
+            "slow (minutes for ~45 frames).  Use CUDA for interactive use."
+        )
+
+    T = video_frames.shape[0]
+    frames = video_frames.to(device).permute(0, 3, 1, 2).to(torch.float32) / 255.0
+    if resize_frames != 1.0:
+        new_h = max(1, int(frames.shape[2] * resize_frames))
+        new_w = max(1, int(frames.shape[3] * resize_frames))
+        frames = F.interpolate(frames, size=(new_h, new_w), mode="area")
+
+    _, _, H, W = frames.shape
+    internal_h = resize_flow * H
+    internal_w = resize_flow * W
+    if internal_h % downscale_factor or internal_w % downscale_factor:
+        logging.warning(
+            "VOIDWarpedNoise: internal noise size %dx%d is not divisible by "
+            "downscale_factor %d; output noise may have artifacts.",
+            internal_h, internal_w, downscale_factor,
+        )
+
+    with torch.no_grad():
+        warper = NoiseWarper(
+            c=noise_channels, h=internal_h, w=internal_w, device=device,
+        )
+        down_h = warper.h // downscale_factor
+        down_w = warper.w // downscale_factor
+        output = torch.empty(
+            (T, down_h, down_w, noise_channels), dtype=torch.float32, device=device,
+        )
+
+        def downscale(noise_chw):
+            # Area-pool to 1/downscale_factor then multiply by downscale_factor
+            # to adjust std (sqrt of pool area == downscale_factor for a
+            # square pool).
+            down = _torch_resize_chw(noise_chw, 1.0 / downscale_factor, "area", copy=False)
+            return down * downscale_factor
+
+        output[0] = downscale(warper.noise).permute(1, 2, 0)
+
+        prev = frames[0]
+        for i in range(1, T):
+            curr = frames[i]
+            flow = raft(prev, curr).to(device)
+            warper(flow[0], flow[1])
+            output[i] = downscale(warper.noise).permute(1, 2, 0)
+            prev = curr
+
+    return output

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:

folder_paths.py

@@ -54,6 +54,8 @@ folder_names_and_paths["audio_encoders"] = ([os.path.join(models_dir, "audio_enc
 
 folder_names_and_paths["frame_interpolation"] = ([os.path.join(models_dir, "frame_interpolation")], supported_pt_extensions)
 
+folder_names_and_paths["optical_flow"] = ([os.path.join(models_dir, "optical_flow")], supported_pt_extensions)
+
 output_directory = os.path.join(base_path, "output")
 temp_directory = os.path.join(base_path, "temp")
 input_directory = os.path.join(base_path, "input")

nodes.py

@@ -958,7 +958,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image", "cogvideox"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),
@@ -968,7 +968,7 @@ class CLIPLoader:
 
     CATEGORY = "advanced/loaders"
 
-    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5\nomnigen2: qwen vl 2.5 3B"
+    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncogvideox: t5 xxl (226-token padding)\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5\nomnigen2: qwen vl 2.5 3B"
 
     def load_clip(self, clip_name, type="stable_diffusion", device="default"):
         clip_type = getattr(comfy.sd.CLIPType, type.upper(), comfy.sd.CLIPType.STABLE_DIFFUSION)
@@ -2445,6 +2445,7 @@ async def init_builtin_extra_nodes():
         "nodes_rtdetr.py",
         "nodes_frame_interpolation.py",
         "nodes_sam3.py",
+        "nodes_void.py",
     ]
 
     import_failed = []