Merge 296b7c7b6d into f5030e26fd

* feat(api-nodes): add HitPaw API nodes

* remove face_soft_2x model as not working

---------

Co-authored-by: Robin Huang <robin.j.huang@gmail.com>

comfy/latent_formats.py

@@ -755,6 +755,10 @@ class ACEAudio(LatentFormat):
     latent_channels = 8
     latent_dimensions = 2
 
+class ACEAudio15(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+
 class ChromaRadiance(LatentFormat):
     latent_channels = 3
     spacial_downscale_ratio = 1

comfy/ldm/kandinsky5/model.py

@@ -6,6 +6,12 @@ import comfy.ldm.common_dit
 from comfy.ldm.modules.attention import optimized_attention
 from comfy.ldm.flux.math import apply_rope1
 from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.kandinsky5.utils_nabla import (
+    fractal_flatten,
+    fractal_unflatten,
+    fast_sta_nabla,
+    nabla,
+)
 
 def attention(q, k, v, heads, transformer_options={}):
     return optimized_attention(
@@ -116,14 +122,17 @@ class SelfAttention(nn.Module):
         result = proj_fn(x).view(*x.shape[:-1], self.num_heads, -1)
         return apply_rope1(norm_fn(result), freqs)
 
-    def _forward(self, x, freqs, transformer_options={}):
+    def _forward(self, x, freqs, sparse_params=None, transformer_options={}):
         q = self._compute_qk(x, freqs, self.to_query, self.query_norm)
         k = self._compute_qk(x, freqs, self.to_key, self.key_norm)
         v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1)
-        out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        if sparse_params is None:
+            out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        else:
+            out = nabla(q, k, v, sparse_params)
         return self.out_layer(out)
 
-    def _forward_chunked(self, x, freqs, transformer_options={}):
+    def _forward_chunked(self, x, freqs, sparse_params=None, transformer_options={}):
         def process_chunks(proj_fn, norm_fn):
             x_chunks = torch.chunk(x, self.num_chunks, dim=1)
             freqs_chunks = torch.chunk(freqs, self.num_chunks, dim=1)
@@ -135,14 +144,17 @@ class SelfAttention(nn.Module):
         q = process_chunks(self.to_query, self.query_norm)
         k = process_chunks(self.to_key, self.key_norm)
         v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1)
-        out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        if sparse_params is None:
+            out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        else:
+            out = nabla(q, k, v, sparse_params)
         return self.out_layer(out)
 
-    def forward(self, x, freqs, transformer_options={}):
+    def forward(self, x, freqs, sparse_params=None, transformer_options={}):
         if x.shape[1] > 8192:
-            return self._forward_chunked(x, freqs, transformer_options=transformer_options)
+            return self._forward_chunked(x, freqs, sparse_params=sparse_params, transformer_options=transformer_options)
         else:
-            return self._forward(x, freqs, transformer_options=transformer_options)
+            return self._forward(x, freqs, sparse_params=sparse_params, transformer_options=transformer_options)
 
 
 class CrossAttention(SelfAttention):
@@ -251,12 +263,12 @@ class TransformerDecoderBlock(nn.Module):
         self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings)
 
-    def forward(self, visual_embed, text_embed, time_embed, freqs, transformer_options={}):
+    def forward(self, visual_embed, text_embed, time_embed, freqs, sparse_params=None, transformer_options={}):
         self_attn_params, cross_attn_params, ff_params = torch.chunk(self.visual_modulation(time_embed), 3, dim=-1)
         # self attention
         shift, scale, gate = get_shift_scale_gate(self_attn_params)
         visual_out = apply_scale_shift_norm(self.self_attention_norm, visual_embed, scale, shift)
-        visual_out = self.self_attention(visual_out, freqs, transformer_options=transformer_options)
+        visual_out = self.self_attention(visual_out, freqs, sparse_params=sparse_params, transformer_options=transformer_options)
         visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
         # cross attention
         shift, scale, gate = get_shift_scale_gate(cross_attn_params)
@@ -369,21 +381,82 @@ class Kandinsky5(nn.Module):
 
         visual_embed = self.visual_embeddings(x)
         visual_shape = visual_embed.shape[:-1]
-        visual_embed = visual_embed.flatten(1, -2)
 
         blocks_replace = patches_replace.get("dit", {})
         transformer_options["total_blocks"] = len(self.visual_transformer_blocks)
         transformer_options["block_type"] = "double"
+
+        B, _, T, H, W = x.shape
+        NABLA_THR = 31 # long (10 sec) generation
+        if T > NABLA_THR:
+            assert self.patch_size[0] == 1
+
+            # pro video model uses lower P at higher resolutions
+            P = 0.7 if self.model_dim == 4096 and H * W >= 14080 else 0.9
+
+            freqs = freqs.view(freqs.shape[0], *visual_shape[1:], *freqs.shape[2:])
+            visual_embed, freqs = fractal_flatten(visual_embed, freqs, visual_shape[1:])
+            pt, ph, pw = self.patch_size
+            T, H, W = T // pt, H // ph, W // pw
+
+            wT, wW, wH = 11, 3, 3
+            sta_mask = fast_sta_nabla(T, H // 8, W // 8, wT, wH, wW, device=x.device)
+
+            sparse_params = dict(
+                sta_mask=sta_mask.unsqueeze_(0).unsqueeze_(0),
+                attention_type="nabla",
+                to_fractal=True,
+                P=P,
+                wT=wT, wW=wW, wH=wH,
+                add_sta=True,
+                visual_shape=(T, H, W),
+                method="topcdf",
+            )
+        else:
+            sparse_params = None
+            visual_embed = visual_embed.flatten(1, -2)
+
         for i, block in enumerate(self.visual_transformer_blocks):
             transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
-                    return block(x=args["x"], context=args["context"], time_embed=args["time_embed"], freqs=args["freqs"], transformer_options=args.get("transformer_options"))
-                visual_embed = blocks_replace[("double_block", i)]({"x": visual_embed, "context": context, "time_embed": time_embed, "freqs": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})["x"]
+                    return block(
+                        x=args["x"],
+                        context=args["context"],
+                        time_embed=args["time_embed"],
+                        freqs=args["freqs"],
+                        sparse_params=args.get("sparse_params"),
+                        transformer_options=args.get("transformer_options"),
+                    )
+                visual_embed = blocks_replace[("double_block", i)](
+                    {
+                        "x": visual_embed,
+                        "context": context,
+                        "time_embed": time_embed,
+                        "freqs": freqs,
+                        "sparse_params": sparse_params,
+                        "transformer_options": transformer_options,
+                    },
+                    {"original_block": block_wrap},
+                )["x"]
             else:
-                visual_embed = block(visual_embed, context, time_embed, freqs=freqs, transformer_options=transformer_options)
+                visual_embed = block(
+                    visual_embed,
+                    context,
+                    time_embed,
+                    freqs=freqs,
+                    sparse_params=sparse_params,
+                    transformer_options=transformer_options,
+                )
+
+        if T > NABLA_THR:
+            visual_embed = fractal_unflatten(
+                visual_embed,
+                visual_shape[1:],
+            )
+        else:
+            visual_embed = visual_embed.reshape(*visual_shape, -1)
 
-        visual_embed = visual_embed.reshape(*visual_shape, -1)
         return self.out_layer(visual_embed, time_embed)
 
     def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):

comfy/ldm/kandinsky5/utils_nabla.py

@@ -0,0 +1,146 @@
+import math
+
+import torch
+from torch import Tensor
+from torch.nn.attention.flex_attention import BlockMask, flex_attention
+
+
+def fractal_flatten(x, rope, shape):
+    pixel_size = 8
+    x = local_patching(x, shape, (1, pixel_size, pixel_size), dim=1)
+    rope = local_patching(rope, shape, (1, pixel_size, pixel_size), dim=1)
+    x = x.flatten(1, 2)
+    rope = rope.flatten(1, 2)
+    return x, rope
+
+
+def fractal_unflatten(x, shape):
+    pixel_size = 8
+    x = x.reshape(x.shape[0], -1, pixel_size**2, x.shape[-1])
+    x = local_merge(x, shape, (1, pixel_size, pixel_size), dim=1)
+    return x
+
+def local_patching(x, shape, group_size, dim=0):
+    duration, height, width = shape
+    g1, g2, g3 = group_size
+    x = x.reshape(
+        *x.shape[:dim],
+        duration // g1,
+        g1,
+        height // g2,
+        g2,
+        width // g3,
+        g3,
+        *x.shape[dim + 3 :]
+    )
+    x = x.permute(
+        *range(len(x.shape[:dim])),
+        dim,
+        dim + 2,
+        dim + 4,
+        dim + 1,
+        dim + 3,
+        dim + 5,
+        *range(dim + 6, len(x.shape))
+    )
+    x = x.flatten(dim, dim + 2).flatten(dim + 1, dim + 3)
+    return x
+
+
+def local_merge(x, shape, group_size, dim=0):
+    duration, height, width = shape
+    g1, g2, g3 = group_size
+    x = x.reshape(
+        *x.shape[:dim],
+        duration // g1,
+        height // g2,
+        width // g3,
+        g1,
+        g2,
+        g3,
+        *x.shape[dim + 2 :]
+    )
+    x = x.permute(
+        *range(len(x.shape[:dim])),
+        dim,
+        dim + 3,
+        dim + 1,
+        dim + 4,
+        dim + 2,
+        dim + 5,
+        *range(dim + 6, len(x.shape))
+    )
+    x = x.flatten(dim, dim + 1).flatten(dim + 1, dim + 2).flatten(dim + 2, dim + 3)
+    return x
+
+def fast_sta_nabla(T: int, H: int, W: int, wT: int = 3, wH: int = 3, wW: int = 3, device="cuda") -> Tensor:
+    l = torch.Tensor([T, H, W]).amax()
+    r = torch.arange(0, l, 1, dtype=torch.int16, device=device)
+    mat = (r.unsqueeze(1) - r.unsqueeze(0)).abs()
+    sta_t, sta_h, sta_w = (
+        mat[:T, :T].flatten(),
+        mat[:H, :H].flatten(),
+        mat[:W, :W].flatten(),
+    )
+    sta_t = sta_t <= wT // 2
+    sta_h = sta_h <= wH // 2
+    sta_w = sta_w <= wW // 2
+    sta_hw = (
+        (sta_h.unsqueeze(1) * sta_w.unsqueeze(0))
+        .reshape(H, H, W, W)
+        .transpose(1, 2)
+        .flatten()
+    )
+    sta = (
+        (sta_t.unsqueeze(1) * sta_hw.unsqueeze(0))
+        .reshape(T, T, H * W, H * W)
+        .transpose(1, 2)
+    )
+    return sta.reshape(T * H * W, T * H * W)
+
+def nablaT_v2(q: Tensor, k: Tensor, sta: Tensor, thr: float = 0.9) -> BlockMask:
+    # Map estimation
+    B, h, S, D = q.shape
+    s1 = S // 64
+    qa = q.reshape(B, h, s1, 64, D).mean(-2)
+    ka = k.reshape(B, h, s1, 64, D).mean(-2).transpose(-2, -1)
+    map = qa @ ka
+
+    map = torch.softmax(map / math.sqrt(D), dim=-1)
+    # Map binarization
+    vals, inds = map.sort(-1)
+    cvals = vals.cumsum_(-1)
+    mask = (cvals >= 1 - thr).int()
+    mask = mask.gather(-1, inds.argsort(-1))
+    mask = torch.logical_or(mask, sta)
+
+    # BlockMask creation
+    kv_nb = mask.sum(-1).to(torch.int32)
+    kv_inds = mask.argsort(dim=-1, descending=True).to(torch.int32)
+    return BlockMask.from_kv_blocks(
+        torch.zeros_like(kv_nb), kv_inds, kv_nb, kv_inds, BLOCK_SIZE=64, mask_mod=None
+    )
+
+@torch.compile(mode="max-autotune-no-cudagraphs", dynamic=True)
+def nabla(query, key, value, sparse_params=None):
+    query = query.transpose(1, 2).contiguous()
+    key = key.transpose(1, 2).contiguous()
+    value = value.transpose(1, 2).contiguous()
+    block_mask = nablaT_v2(
+        query,
+        key,
+        sparse_params["sta_mask"],
+        thr=sparse_params["P"],
+    )
+    out = (
+        flex_attention(
+            query,
+            key,
+            value,
+            block_mask=block_mask
+        )
+        .transpose(1, 2)
+        .contiguous()
+    )
+    out = out.flatten(-2, -1)
+    return out

comfy/model_base.py

@@ -50,6 +50,7 @@ import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
+import comfy.ldm.ace.ace_step15
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -1540,6 +1541,47 @@ class ACEStep(BaseModel):
         out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
         return out
 
+class ACEStep15(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.ace_step15.AceStepConditionGenerationModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        device = kwargs["device"]
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_embed'] = comfy.conds.CONDRegular(conditioning_lyrics)
+
+        refer_audio = kwargs.get("reference_audio_timbre_latents", None)
+        if refer_audio is None or len(refer_audio) == 0:
+            refer_audio = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
+                                        2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
+                                        -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
+                                        7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
+                                        2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
+                                        -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
+                                        5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
+                                        -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
+                                        2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
+                                        1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
+                                        -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
+                                        -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
+                                        7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, 750)
+        else:
+            refer_audio = refer_audio[-1]
+        out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
+
+        audio_codes = kwargs.get("audio_codes", None)
+        if audio_codes is not None:
+            out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+
+        return out
+
 class Omnigen2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)

comfy/model_detection.py

@@ -655,6 +655,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
+    if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys:
+        dit_config = {}
+        dit_config["audio_model"] = "ace1.5"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/model_management.py

@@ -19,7 +19,8 @@
 import psutil
 import logging
 from enum import Enum
-from comfy.cli_args import args, PerformanceFeature
+from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
+import threading
 import torch
 import sys
 import platform
@@ -650,7 +651,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 soft_empty_cache()
     return unloaded_models
 
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     cleanup_models_gc()
     global vram_state
 
@@ -746,6 +747,26 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         current_loaded_models.insert(0, loaded_model)
     return
 
+def load_models_gpu_thread(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load):
+    with torch.inference_mode():
+        load_models_gpu_orig(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
+        soft_empty_cache()
+
+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+    #Deliberately load models outside of the Aimdo mempool so they can be retained accross
+    #nodes. Use a dummy thread to do it as pytorch documents that mempool contexts are
+    #thread local. So exploit that to escape context
+    if enables_dynamic_vram():
+        t = threading.Thread(
+            target=load_models_gpu_thread,
+            args=(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
+        )
+        t.start()
+        t.join()
+    else:
+        load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights,
+                             minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
+
 def load_model_gpu(model):
     return load_models_gpu([model])
 
@@ -1112,11 +1133,11 @@ def get_cast_buffer(offload_stream, device, size, ref):
             return None
         if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2):
             #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now
-            torch.cuda.synchronize()
+            synchronize()
             del STREAM_CAST_BUFFERS[offload_stream]
             del cast_buffer
             #FIXME: This doesn't work in Aimdo because mempool cant clear cache
-            torch.cuda.empty_cache()
+            soft_empty_cache()
         with wf_context:
             cast_buffer = torch.empty((size), dtype=torch.int8, device=device)
             STREAM_CAST_BUFFERS[offload_stream] = cast_buffer
@@ -1132,9 +1153,7 @@ def reset_cast_buffers():
     for offload_stream in STREAM_CAST_BUFFERS:
         offload_stream.synchronize()
     STREAM_CAST_BUFFERS.clear()
-    if comfy.memory_management.aimdo_allocator is None:
-        #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist
-        torch.cuda.empty_cache()
+    soft_empty_cache()
 
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
@@ -1284,7 +1303,7 @@ def discard_cuda_async_error():
         a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         _ = a + b
-        torch.cuda.synchronize()
+        synchronize()
     except torch.AcceleratorError:
         #Dump it! We already know about it from the synchronous return
         pass
@@ -1688,6 +1707,12 @@ def lora_compute_dtype(device):
     LORA_COMPUTE_DTYPES[device] = dtype
     return dtype
 
+def synchronize():
+    if is_intel_xpu():
+        torch.xpu.synchronize()
+    elif torch.cuda.is_available():
+        torch.cuda.synchronize()
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
@@ -1713,9 +1738,6 @@ def debug_memory_summary():
         return torch.cuda.memory.memory_summary()
     return ""
 
-#TODO: might be cleaner to put this somewhere else
-import threading
-
 class InterruptProcessingException(Exception):
     pass
 

comfy/model_patcher.py

@@ -1597,7 +1597,7 @@ class ModelPatcherDynamic(ModelPatcher):
 
         if unpatch_weights:
             self.partially_unload_ram(1e32)
-            self.partially_unload(None)
+            self.partially_unload(None, 1e32)
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
         assert not force_patch_weights #See above

comfy/sd.py

@@ -59,6 +59,7 @@ import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.jina_clip_2
 import comfy.text_encoders.newbie
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 import comfy.model_patcher
 import comfy.lora
@@ -452,6 +453,8 @@ class VAE:
         self.extra_1d_channel = None
         self.crop_input = True
 
+        self.audio_sample_rate = 44100
+
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
                 encoder_config = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
@@ -549,14 +552,25 @@ class VAE:
                                                                     encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
                                                                     decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
             elif "decoder.layers.1.layers.0.beta" in sd:
-                self.first_stage_model = AudioOobleckVAE()
+                config = {}
+                param_key = None
+                if "decoder.layers.2.layers.1.weight_v" in sd:
+                    param_key = "decoder.layers.2.layers.1.weight_v"
+                if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd:
+                    param_key = "decoder.layers.2.layers.1.parametrizations.weight.original1"
+                if param_key is not None:
+                    if sd[param_key].shape[-1] == 12:
+                        config["strides"] = [2, 4, 4, 6, 10]
+                        self.audio_sample_rate = 48000
+
+                self.first_stage_model = AudioOobleckVAE(**config)
                 self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype)
                 self.latent_channels = 64
                 self.output_channels = 2
                 self.pad_channel_value = "replicate"
                 self.upscale_ratio = 2048
-                self.downscale_ratio =  2048
+                self.downscale_ratio = 2048
                 self.latent_dim = 1
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
@@ -1427,6 +1441,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_data_jina = clip_data[0]
             tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
             tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
+        elif clip_type == CLIPType.ACE:
+            clip_target.clip = comfy.text_encoders.ace15.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer

comfy/sd1_clip.py

@@ -155,6 +155,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.execution_device = options.get("execution_device", self.execution_device)
         if isinstance(self.layer, list) or self.layer == "all":
             pass
+        elif isinstance(layer_idx, list):
+            self.layer = layer_idx
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:

comfy/supported_models.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 from . import supported_models_base
 from . import latent_formats
@@ -1596,6 +1597,38 @@ class Kandinsky5Image(Kandinsky5):
         return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
+class ACEStep15(supported_models_base.BASE):
+    unet_config = {
+        "audio_model": "ace1.5",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 3.0,
+    }
+
+    latent_format = comfy.latent_formats.ACEAudio15
+
+    memory_usage_factor = 4.7
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.ACEStep15(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**hunyuan_detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace15.py

@@ -0,0 +1,222 @@
+from .anima import Qwen3Tokenizer
+import comfy.text_encoders.llama
+from comfy import sd1_clip
+import torch
+import math
+import comfy.utils
+
+
+def sample_manual_loop_no_classes(
+    model,
+    ids=None,
+    paddings=[],
+    execution_dtype=None,
+    cfg_scale: float = 2.0,
+    temperature: float = 0.85,
+    top_p: float = 0.9,
+    top_k: int = None,
+    seed: int = 1,
+    min_tokens: int = 1,
+    max_new_tokens: int = 2048,
+    audio_start_id: int = 151669,  # The cutoff ID for audio codes
+    eos_token_id: int = 151645,
+):
+    device = model.execution_device
+
+    if execution_dtype is None:
+        if comfy.model_management.should_use_bf16(device):
+            execution_dtype = torch.bfloat16
+        else:
+            execution_dtype = torch.float32
+
+    embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device)
+    for i, t in enumerate(paddings):
+        attention_mask[i, :t] = 0
+        attention_mask[i, t:] = 1
+
+    output_audio_codes = []
+    past_key_values = []
+    generator = torch.Generator(device=device)
+    generator.manual_seed(seed)
+    model_config = model.transformer.model.config
+
+    for x in range(model_config.num_hidden_layers):
+        past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0))
+
+    progress_bar = comfy.utils.ProgressBar(max_new_tokens)
+
+    for step in range(max_new_tokens):
+        outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values)
+        next_token_logits = model.transformer.logits(outputs[0])[:, -1]
+        past_key_values = outputs[2]
+
+        cond_logits = next_token_logits[0:1]
+        uncond_logits = next_token_logits[1:2]
+        cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits)
+
+        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+            eos_score = cfg_logits[:, eos_token_id].clone()
+
+        # Only generate audio tokens
+        cfg_logits[:, :audio_start_id] = float('-inf')
+
+        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+            cfg_logits[:, eos_token_id] = eos_score
+
+        if top_k is not None and top_k > 0:
+            top_k_vals, _ = torch.topk(cfg_logits, top_k)
+            min_val = top_k_vals[..., -1, None]
+            cfg_logits[cfg_logits < min_val] = float('-inf')
+
+        if top_p is not None and top_p < 1.0:
+            sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True)
+            cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+            cfg_logits[indices_to_remove] = float('-inf')
+
+        if temperature > 0:
+            cfg_logits = cfg_logits / temperature
+            next_token = torch.multinomial(torch.softmax(cfg_logits, dim=-1), num_samples=1, generator=generator).squeeze(1)
+        else:
+            next_token = torch.argmax(cfg_logits, dim=-1)
+
+        token = next_token.item()
+
+        if token == eos_token_id:
+            break
+
+        embed, _, _, _ = model.process_tokens([[token]], device)
+        embeds = embed.repeat(2, 1, 1)
+        attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1)
+
+        output_audio_codes.append(token - audio_start_id)
+        progress_bar.update_absolute(step)
+
+    return output_audio_codes
+
+
+def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0):
+    cfg_scale = 2.0
+
+    positive = [[token for token, _ in inner_list] for inner_list in positive]
+    negative = [[token for token, _ in inner_list] for inner_list in negative]
+    positive = positive[0]
+    negative = negative[0]
+
+    neg_pad = 0
+    if len(negative) < len(positive):
+        neg_pad = (len(positive) - len(negative))
+        negative = [model.special_tokens["pad"]] * neg_pad + negative
+
+    pos_pad = 0
+    if len(negative) > len(positive):
+        pos_pad = (len(negative) - len(positive))
+        positive = [model.special_tokens["pad"]] * pos_pad + positive
+
+    paddings = [pos_pad, neg_pad]
+    return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
+
+
+class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
+        out = {}
+        lyrics = kwargs.get("lyrics", "")
+        bpm = kwargs.get("bpm", 120)
+        duration = kwargs.get("duration", 120)
+        keyscale = kwargs.get("keyscale", "C major")
+        timesignature = kwargs.get("timesignature", 2)
+        language = kwargs.get("language", "en")
+        seed = kwargs.get("seed", 0)
+
+        duration = math.ceil(duration)
+        meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature)
+        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n<think>\n{}\n</think>\n\n<|im_end|>\n"
+
+        meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration)
+        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True)
+        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True)
+
+        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs)
+        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
+        out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed}
+        return out
+
+
+class Qwen3_06BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_06B_ACE15, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Qwen3_2B_ACE15(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class ACE15TEModel(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None, dtype_llama=None, model_options={}):
+        super().__init__()
+        if dtype_llama is None:
+            dtype_llama = dtype
+
+        self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options)
+        self.qwen3_2b = Qwen3_2B_ACE15(device=device, dtype=dtype_llama, model_options=model_options)
+        self.dtypes = set([dtype, dtype_llama])
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_base = token_weight_pairs["qwen3_06b"]
+        token_weight_pairs_lyrics = token_weight_pairs["lyrics"]
+
+        self.qwen3_06b.set_clip_options({"layer": None})
+        base_out, _, extra = self.qwen3_06b.encode_token_weights(token_weight_pairs_base)
+        self.qwen3_06b.set_clip_options({"layer": [0]})
+        lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics)
+
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        audio_codes = generate_audio_codes(self.qwen3_2b, token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"])
+
+        return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]}
+
+    def set_clip_options(self, options):
+        self.qwen3_06b.set_clip_options(options)
+        self.qwen3_2b.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.qwen3_06b.reset_clip_options()
+        self.qwen3_2b.reset_clip_options()
+
+    def load_sd(self, sd):
+        if "model.layers.0.post_attention_layernorm.weight" in sd:
+            shape = sd["model.layers.0.post_attention_layernorm.weight"].shape
+            if shape[0] == 1024:
+                return self.qwen3_06b.load_sd(sd)
+            else:
+                return self.qwen3_2b.load_sd(sd)
+
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        constant = 0.4375
+        if comfy.model_management.should_use_bf16(device):
+            constant *= 0.5
+
+        token_weight_pairs = token_weight_pairs.get("lm_prompt", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        num_tokens += lm_metadata['min_tokens']
+        return num_tokens * constant * 1024 * 1024
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class ACE15TEModel_(ACE15TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype_llama=dtype_llama, dtype=dtype, model_options=model_options)
+    return ACE15TEModel_

comfy/text_encoders/llama.py

@@ -103,6 +103,52 @@ class Qwen3_06BConfig:
     final_norm: bool = True
     lm_head: bool = False
 
+@dataclass
+class Qwen3_06B_ACE15_Config:
+    vocab_size: int = 151669
+    hidden_size: int = 1024
+    intermediate_size: int = 3072
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 32768
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
+@dataclass
+class Qwen3_2B_ACE15_lm_Config:
+    vocab_size: int = 217204
+    hidden_size: int = 2048
+    intermediate_size: int = 6144
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
 @dataclass
 class Qwen3_4BConfig:
     vocab_size: int = 151936
@@ -729,6 +775,27 @@ class Qwen3_06B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_06B_ACE15_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_2B_ACE15_lm_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+    def logits(self, x):
+        return torch.nn.functional.linear(x[:, -1:], self.model.embed_tokens.weight.to(x), None)
+
 class Qwen3_4B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy_api_nodes/apis/hitpaw.py

@@ -0,0 +1,51 @@
+from typing import TypedDict
+
+from pydantic import BaseModel, Field
+
+
+class InputVideoModel(TypedDict):
+    model: str
+    resolution: str
+
+
+class ImageEnhanceTaskCreateRequest(BaseModel):
+    model_name: str = Field(...)
+    img_url: str = Field(...)
+    extension: str = Field(".png")
+    exif: bool = Field(False)
+    DPI: int | None = Field(None)
+
+
+class VideoEnhanceTaskCreateRequest(BaseModel):
+    video_url: str = Field(...)
+    extension: str = Field(".mp4")
+    model_name: str | None = Field(...)
+    resolution: list[int] = Field(..., description="Target resolution [width, height]")
+    original_resolution: list[int] = Field(..., description="Original video resolution [width, height]")
+
+
+class TaskCreateDataResponse(BaseModel):
+    job_id: str = Field(...)
+    consume_coins: int | None = Field(None)
+
+
+class TaskStatusPollRequest(BaseModel):
+    job_id: str = Field(...)
+
+
+class TaskCreateResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskCreateDataResponse | None = Field(None)
+
+
+class TaskStatusDataResponse(BaseModel):
+    job_id: str = Field(...)
+    status: str = Field(...)
+    res_url: str = Field("")
+
+
+class TaskStatusResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskStatusDataResponse = Field(...)

comfy_api_nodes/nodes_hitpaw.py

@@ -0,0 +1,342 @@
+import math
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.hitpaw import (
+    ImageEnhanceTaskCreateRequest,
+    InputVideoModel,
+    TaskCreateDataResponse,
+    TaskCreateResponse,
+    TaskStatusPollRequest,
+    TaskStatusResponse,
+    VideoEnhanceTaskCreateRequest,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    download_url_to_image_tensor,
+    download_url_to_video_output,
+    downscale_image_tensor,
+    get_image_dimensions,
+    poll_op,
+    sync_op,
+    upload_image_to_comfyapi,
+    upload_video_to_comfyapi,
+    validate_video_duration,
+)
+
+VIDEO_MODELS_MODELS_MAP = {
+    "Portrait Restore Model (1x)": "portrait_restore_1x",
+    "Portrait Restore Model (2x)": "portrait_restore_2x",
+    "General Restore Model (1x)": "general_restore_1x",
+    "General Restore Model (2x)": "general_restore_2x",
+    "General Restore Model (4x)": "general_restore_4x",
+    "Ultra HD Model (2x)": "ultrahd_restore_2x",
+    "Generative Model (1x)": "generative_1x",
+}
+
+# Resolution name to target dimension (shorter side) in pixels
+RESOLUTION_TARGET_MAP = {
+    "720p": 720,
+    "1080p": 1080,
+    "2K/QHD": 1440,
+    "4K/UHD": 2160,
+    "8K": 4320,
+}
+
+# Square (1:1) resolutions use standard square dimensions
+RESOLUTION_SQUARE_MAP = {
+    "720p": 720,
+    "1080p": 1080,
+    "2K/QHD": 1440,
+    "4K/UHD": 2048,  # DCI 4K square
+    "8K": 4096,  # DCI 8K square
+}
+
+# Models with limited resolution support (no 8K)
+LIMITED_RESOLUTION_MODELS = {"Generative Model (1x)"}
+
+# Resolution options for different model types
+RESOLUTIONS_LIMITED = ["original", "720p", "1080p", "2K/QHD", "4K/UHD"]
+RESOLUTIONS_FULL = ["original", "720p", "1080p", "2K/QHD", "4K/UHD", "8K"]
+
+# Maximum output resolution in pixels
+MAX_PIXELS_GENERATIVE = 32_000_000
+MAX_MP_GENERATIVE = MAX_PIXELS_GENERATIVE // 1_000_000
+
+
+class HitPawGeneralImageEnhance(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HitPawGeneralImageEnhance",
+            display_name="HitPaw General Image Enhance",
+            category="api node/image/HitPaw",
+            description="Upscale low-resolution images to super-resolution, eliminate artifacts and noise. "
+            f"Maximum output: {MAX_MP_GENERATIVE} megapixels.",
+            inputs=[
+                IO.Combo.Input("model", options=["generative_portrait", "generative"]),
+                IO.Image.Input("image"),
+                IO.Combo.Input("upscale_factor", options=[1, 2, 4]),
+                IO.Boolean.Input(
+                    "auto_downscale",
+                    default=False,
+                    tooltip="Automatically downscale input image if output would exceed the limit.",
+                ),
+            ],
+            outputs=[
+                IO.Image.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=["model"]),
+                expr="""
+                (
+                  $prices := {
+                    "generative_portrait": {"min": 0.02, "max": 0.06},
+                    "generative": {"min": 0.05, "max": 0.15}
+                  };
+                  $price := $lookup($prices, widgets.model);
+                  {
+                    "type": "range_usd",
+                    "min_usd": $price.min,
+                    "max_usd": $price.max
+                  }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: Input.Image,
+        upscale_factor: int,
+        auto_downscale: bool,
+    ) -> IO.NodeOutput:
+        height, width = get_image_dimensions(image)
+        requested_scale = upscale_factor
+        output_pixels = height * width * requested_scale * requested_scale
+        if output_pixels > MAX_PIXELS_GENERATIVE:
+            if auto_downscale:
+                input_pixels = width * height
+                scale = 1
+                max_input_pixels = MAX_PIXELS_GENERATIVE
+
+                for candidate in [4, 2, 1]:
+                    if candidate > requested_scale:
+                        continue
+                    scale_output_pixels = input_pixels * candidate * candidate
+                    if scale_output_pixels <= MAX_PIXELS_GENERATIVE:
+                        scale = candidate
+                        max_input_pixels = None
+                        break
+                    # Check if we can downscale input by at most 2x to fit
+                    downscale_ratio = math.sqrt(scale_output_pixels / MAX_PIXELS_GENERATIVE)
+                    if downscale_ratio <= 2.0:
+                        scale = candidate
+                        max_input_pixels = MAX_PIXELS_GENERATIVE // (candidate * candidate)
+                        break
+
+                if max_input_pixels is not None:
+                    image = downscale_image_tensor(image, total_pixels=max_input_pixels)
+                upscale_factor = scale
+            else:
+                output_width = width * requested_scale
+                output_height = height * requested_scale
+                raise ValueError(
+                    f"Output size ({output_width}x{output_height} = {output_pixels:,} pixels) "
+                    f"exceeds maximum allowed size of {MAX_PIXELS_GENERATIVE:,} pixels ({MAX_MP_GENERATIVE}MP). "
+                    f"Enable auto_downscale or use a smaller input image or a lower upscale factor."
+                )
+
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/photo-enhancer", method="POST"),
+            response_model=TaskCreateResponse,
+            data=ImageEnhanceTaskCreateRequest(
+                model_name=f"{model}_{upscale_factor}x",
+                img_url=await upload_image_to_comfyapi(cls, image, total_pixels=None),
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}")
+        request_price = initial_res.data.consume_coins / 1000
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"),
+            data=TaskCreateDataResponse(job_id=initial_res.data.job_id),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda x: x.data.status,
+            price_extractor=lambda x: request_price,
+            poll_interval=10.0,
+            max_poll_attempts=480,
+        )
+        return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.res_url))
+
+
+class HitPawVideoEnhance(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        model_options = []
+        for model_name in VIDEO_MODELS_MODELS_MAP:
+            if model_name in LIMITED_RESOLUTION_MODELS:
+                resolutions = RESOLUTIONS_LIMITED
+            else:
+                resolutions = RESOLUTIONS_FULL
+            model_options.append(
+                IO.DynamicCombo.Option(
+                    model_name,
+                    [IO.Combo.Input("resolution", options=resolutions)],
+                )
+            )
+
+        return IO.Schema(
+            node_id="HitPawVideoEnhance",
+            display_name="HitPaw Video Enhance",
+            category="api node/video/HitPaw",
+            description="Upscale low-resolution videos to high resolution, eliminate artifacts and noise. "
+            "Prices shown are per second of video.",
+            inputs=[
+                IO.DynamicCombo.Input("model", options=model_options),
+                IO.Video.Input("video"),
+            ],
+            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=["model", "model.resolution"]),
+                expr="""
+                (
+                  $m := $lookup(widgets, "model");
+                  $res := $lookup(widgets, "model.resolution");
+                  $standard_model_prices := {
+                    "original": {"min": 0.01, "max": 0.198},
+                    "720p": {"min": 0.01, "max": 0.06},
+                    "1080p": {"min": 0.015, "max": 0.09},
+                    "2k/qhd": {"min": 0.02, "max": 0.117},
+                    "4k/uhd": {"min": 0.025, "max": 0.152},
+                    "8k": {"min": 0.033, "max": 0.198}
+                  };
+                  $ultra_hd_model_prices := {
+                    "original": {"min": 0.015, "max": 0.264},
+                    "720p": {"min": 0.015, "max": 0.092},
+                    "1080p": {"min": 0.02, "max": 0.12},
+                    "2k/qhd": {"min": 0.026, "max": 0.156},
+                    "4k/uhd": {"min": 0.034, "max": 0.203},
+                    "8k": {"min": 0.044, "max": 0.264}
+                  };
+                  $generative_model_prices := {
+                    "original": {"min": 0.015, "max": 0.338},
+                    "720p": {"min": 0.008, "max": 0.090},
+                    "1080p": {"min": 0.05, "max": 0.15},
+                    "2k/qhd": {"min": 0.038, "max": 0.225},
+                    "4k/uhd": {"min": 0.056, "max": 0.338}
+                  };
+                  $prices := $contains($m, "ultra hd") ? $ultra_hd_model_prices :
+                             $contains($m, "generative") ? $generative_model_prices :
+                             $standard_model_prices;
+                  $price := $lookup($prices, $res);
+                  {
+                    "type": "range_usd",
+                    "min_usd": $price.min,
+                    "max_usd": $price.max,
+                    "format": {"approximate": true, "suffix": "/second"}
+                  }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: InputVideoModel,
+        video: Input.Video,
+    ) -> IO.NodeOutput:
+        validate_video_duration(video, min_duration=0.5, max_duration=60 * 60)
+        resolution = model["resolution"]
+        src_width, src_height = video.get_dimensions()
+
+        if resolution == "original":
+            output_width = src_width
+            output_height = src_height
+        else:
+            if src_width == src_height:
+                target_size = RESOLUTION_SQUARE_MAP[resolution]
+                if target_size < src_width:
+                    raise ValueError(
+                        f"Selected resolution {resolution} ({target_size}x{target_size}) is smaller than "
+                        f"the input video ({src_width}x{src_height}). Please select a higher resolution or 'original'."
+                    )
+                output_width = target_size
+                output_height = target_size
+            else:
+                min_dimension = min(src_width, src_height)
+                target_size = RESOLUTION_TARGET_MAP[resolution]
+                if target_size < min_dimension:
+                    raise ValueError(
+                        f"Selected resolution {resolution} ({target_size}p) is smaller than "
+                        f"the input video's shorter dimension ({min_dimension}p). "
+                        f"Please select a higher resolution or 'original'."
+                    )
+                if src_width > src_height:
+                    output_height = target_size
+                    output_width = int(target_size * (src_width / src_height))
+                else:
+                    output_width = target_size
+                    output_height = int(target_size * (src_height / src_width))
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/video-enhancer", method="POST"),
+            response_model=TaskCreateResponse,
+            data=VideoEnhanceTaskCreateRequest(
+                video_url=await upload_video_to_comfyapi(cls, video),
+                resolution=[output_width, output_height],
+                original_resolution=[src_width, src_height],
+                model_name=VIDEO_MODELS_MODELS_MAP[model["model"]],
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        request_price = initial_res.data.consume_coins / 1000
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}")
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"),
+            data=TaskStatusPollRequest(job_id=initial_res.data.job_id),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda x: x.data.status,
+            price_extractor=lambda x: request_price,
+            poll_interval=10.0,
+            max_poll_attempts=320,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.data.res_url))
+
+
+class HitPawExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            HitPawGeneralImageEnhance,
+            HitPawVideoEnhance,
+        ]
+
+
+async def comfy_entrypoint() -> HitPawExtension:
+    return HitPawExtension()

comfy_api_nodes/util/upload_helpers.py

@@ -94,7 +94,7 @@ async def upload_image_to_comfyapi(
     *,
     mime_type: str | None = None,
     wait_label: str | None = "Uploading",
-    total_pixels: int = 2048 * 2048,
+    total_pixels: int | None = 2048 * 2048,
 ) -> str:
     """Uploads a single image to ComfyUI API and returns its download URL."""
     return (

comfy_extras/nodes_ace.py

@@ -28,12 +28,39 @@ class TextEncodeAceStepAudio(io.ComfyNode):
         conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength})
         return io.NodeOutput(conditioning)
 
+class TextEncodeAceStepAudio15(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncodeAceStepAudio1.5",
+            category="conditioning",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.String.Input("tags", multiline=True, dynamic_prompts=True),
+                io.String.Input("lyrics", multiline=True, dynamic_prompts=True),
+                io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True),
+                io.Int.Input("bpm", default=120, min=10, max=300),
+                io.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1),
+                io.Combo.Input("timesignature", options=['2', '3', '4', '6']),
+                io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
+                io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
+            ],
+            outputs=[io.Conditioning.Output()],
+        )
+
+    @classmethod
+    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale) -> io.NodeOutput:
+        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed)
+        conditioning = clip.encode_from_tokens_scheduled(tokens)
+        return io.NodeOutput(conditioning)
+
 
 class EmptyAceStepLatentAudio(io.ComfyNode):
     @classmethod
     def define_schema(cls):
         return io.Schema(
             node_id="EmptyAceStepLatentAudio",
+            display_name="Empty Ace Step 1.0 Latent Audio",
             category="latent/audio",
             inputs=[
                 io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1),
@@ -51,12 +78,60 @@ class EmptyAceStepLatentAudio(io.ComfyNode):
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
 
+class EmptyAceStep15LatentAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyAceStep1.5LatentAudio",
+            display_name="Empty Ace Step 1.5 Latent Audio",
+            category="latent/audio",
+            inputs=[
+                io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.01),
+                io.Int.Input(
+                    "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."
+                ),
+            ],
+            outputs=[io.Latent.Output()],
+        )
+
+    @classmethod
+    def execute(cls, seconds, batch_size) -> io.NodeOutput:
+        length = round((seconds * 48000 / 1920))
+        latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent, "type": "audio"})
+
+class ReferenceTimbreAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ReferenceTimbreAudio",
+            category="advanced/conditioning/audio",
+            is_experimental=True,
+            description="This node sets the reference audio for timbre (for ace step 1.5)",
+            inputs=[
+                io.Conditioning.Input("conditioning"),
+                io.Latent.Input("latent", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, conditioning, latent=None) -> io.NodeOutput:
+        if latent is not None:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_audio_timbre_latents": [latent["samples"]]}, append=True)
+        return io.NodeOutput(conditioning)
+
 class AceExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             TextEncodeAceStepAudio,
             EmptyAceStepLatentAudio,
+            TextEncodeAceStepAudio15,
+            EmptyAceStep15LatentAudio,
+            ReferenceTimbreAudio,
         ]
 
 async def comfy_entrypoint() -> AceExtension:

comfy_extras/nodes_audio.py

@@ -82,13 +82,14 @@ class VAEEncodeAudio(IO.ComfyNode):
     @classmethod
     def execute(cls, vae, audio) -> IO.NodeOutput:
         sample_rate = audio["sample_rate"]
-        if 44100 != sample_rate:
-            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, 44100)
+        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+        if vae_sample_rate != sample_rate:
+            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, vae_sample_rate)
         else:
             waveform = audio["waveform"]
 
         t = vae.encode(waveform.movedim(1, -1))
-        return IO.NodeOutput({"samples":t})
+        return IO.NodeOutput({"samples": t})
 
     encode = execute  # TODO: remove
 
@@ -114,7 +115,8 @@ class VAEDecodeAudio(IO.ComfyNode):
         std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0
         std[std < 1.0] = 1.0
         audio /= std
-        return IO.NodeOutput({"waveform": audio, "sample_rate": 44100 if "sample_rate" not in samples else samples["sample_rate"]})
+        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+        return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]})
 
     decode = execute  # TODO: remove
 

comfy_extras/nodes_kandinsky5.py

@@ -34,6 +34,9 @@ class Kandinsky5ImageToVideo(io.ComfyNode):
 
     @classmethod
     def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None) -> io.NodeOutput:
+        if length > 121: # 10 sec generation, for nabla
+            height = 128 * round(height / 128)
+            width = 128 * round(width / 128)
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         cond_latent_out = {}
         if start_image is not None:

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.11.1"
+__version__ = "0.12.0"

nodes.py

@@ -1001,7 +1001,7 @@ class DualCLIPLoader:
     def INPUT_TYPES(s):
         return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ),
                               "clip_name2": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie"], ),
+                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie", "ace"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.11.1"
+version = "0.12.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.37.11
-comfyui-workflow-templates==0.8.27
+comfyui-workflow-templates==0.8.31
 comfyui-embedded-docs==0.4.0
 torch
 torchsde