[pull] master from Comfy-Org:master

comfy/context_windows.py

@@ -93,6 +93,50 @@ def init_callbacks(self):
         return {}
 
 
+def slice_cond(cond_value, window: IndexListContextWindow, x_in: torch.Tensor, device, temporal_dim: int, temporal_scale: int=1, temporal_offset: int=0, retain_index_list: list[int]=[]):
+    if not (hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor)):
+        return None
+    cond_tensor = cond_value.cond
+    if temporal_dim >= cond_tensor.ndim:
+        return None
+
+    cond_size = cond_tensor.size(temporal_dim)
+
+    if temporal_scale == 1:
+        expected_size = x_in.size(window.dim) - temporal_offset
+        if cond_size != expected_size:
+            return None
+
+    if temporal_offset == 0 and temporal_scale == 1:
+        sliced = window.get_tensor(cond_tensor, device, dim=temporal_dim, retain_index_list=retain_index_list)
+        return cond_value._copy_with(sliced)
+
+    # skip leading latent positions that have no corresponding conditioning (e.g. reference frames)
+    if temporal_offset > 0:
+        indices = [i - temporal_offset for i in window.index_list[temporal_offset:]]
+        indices = [i for i in indices if 0 <= i]
+    else:
+        indices = list(window.index_list)
+
+    if not indices:
+        return None
+
+    if temporal_scale > 1:
+        scaled = []
+        for i in indices:
+            for k in range(temporal_scale):
+                si = i * temporal_scale + k
+                if si < cond_size:
+                    scaled.append(si)
+        indices = scaled
+        if not indices:
+            return None
+
+    idx = tuple([slice(None)] * temporal_dim + [indices])
+    sliced = cond_tensor[idx].to(device)
+    return cond_value._copy_with(sliced)
+
+
 @dataclass
 class ContextSchedule:
     name: str
@@ -177,10 +221,17 @@ def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: Inde
                                     new_cond_item[cond_key] = result
                                     handled = True
                                     break
+                            if not handled and self._model is not None:
+                                result = self._model.resize_cond_for_context_window(
+                                    cond_key, cond_value, window, x_in, device,
+                                    retain_index_list=self.cond_retain_index_list)
+                                if result is not None:
+                                    new_cond_item[cond_key] = result
+                                    handled = True
                             if handled:
                                 continue
                             if isinstance(cond_value, torch.Tensor):
-                                if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \
+                                if (self.dim < cond_value.ndim and cond_value.size(self.dim) == x_in.size(self.dim)) or \
                                    (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)):
                                     new_cond_item[cond_key] = window.get_tensor(cond_value, device)
                             # Handle audio_embed (temporal dim is 1)
@@ -224,6 +275,7 @@ def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_option
         return context_windows
 
     def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        self._model = model
         self.set_step(timestep, model_options)
         context_windows = self.get_context_windows(model, x_in, model_options)
         enumerated_context_windows = list(enumerate(context_windows))

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -536,6 +536,53 @@ def decode_output_shape(self, input_shape):
         c, (ts, hs, ws), to = self._output_scale
         return (input_shape[0], c, input_shape[2] * ts - to, input_shape[3] * hs, input_shape[4] * ws)
 
+    def run_up(self, idx, sample_ref, ended, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size):
+        sample = sample_ref[0]
+        sample_ref[0] = None
+        if idx >= len(self.up_blocks):
+            sample = self.conv_norm_out(sample)
+            if timestep_shift_scale is not None:
+                shift, scale = timestep_shift_scale
+                sample = sample * (1 + scale) + shift
+            sample = self.conv_act(sample)
+            if ended:
+                mark_conv3d_ended(self.conv_out)
+            sample = self.conv_out(sample, causal=self.causal)
+            if sample is not None and sample.shape[2] > 0:
+                sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
+                t = sample.shape[2]
+                output_buffer[:, :, output_offset[0]:output_offset[0] + t].copy_(sample)
+                output_offset[0] += t
+            return
+
+        up_block = self.up_blocks[idx]
+        if ended:
+            mark_conv3d_ended(up_block)
+        if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
+            sample = checkpoint_fn(up_block)(
+                sample, causal=self.causal, timestep=scaled_timestep
+            )
+        else:
+            sample = checkpoint_fn(up_block)(sample, causal=self.causal)
+
+        if sample is None or sample.shape[2] == 0:
+            return
+
+        total_bytes = sample.numel() * sample.element_size()
+        num_chunks = (total_bytes + max_chunk_size - 1) // max_chunk_size
+
+        if num_chunks == 1:
+            # when we are not chunking, detach our x so the callee can free it as soon as they are done
+            next_sample_ref = [sample]
+            del sample
+            self.run_up(idx + 1, next_sample_ref, ended, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
+            return
+        else:
+            samples = torch.chunk(sample, chunks=num_chunks, dim=2)
+
+            for chunk_idx, sample1 in enumerate(samples):
+                self.run_up(idx + 1, [sample1], ended and chunk_idx == len(samples) - 1, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
+
     def forward_orig(
         self,
         sample: torch.FloatTensor,
@@ -555,6 +602,7 @@ def forward_orig(
         )
 
         timestep_shift_scale = None
+        scaled_timestep = None
         if self.timestep_conditioning:
             assert (
                 timestep is not None
@@ -591,54 +639,7 @@ def forward_orig(
 
         max_chunk_size = get_max_chunk_size(sample.device)
 
-        def run_up(idx, sample_ref, ended):
-            sample = sample_ref[0]
-            sample_ref[0] = None
-            if idx >= len(self.up_blocks):
-                sample = self.conv_norm_out(sample)
-                if timestep_shift_scale is not None:
-                    shift, scale = timestep_shift_scale
-                    sample = sample * (1 + scale) + shift
-                sample = self.conv_act(sample)
-                if ended:
-                    mark_conv3d_ended(self.conv_out)
-                sample = self.conv_out(sample, causal=self.causal)
-                if sample is not None and sample.shape[2] > 0:
-                    sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
-                    t = sample.shape[2]
-                    output_buffer[:, :, output_offset[0]:output_offset[0] + t].copy_(sample)
-                    output_offset[0] += t
-                return
-
-            up_block = self.up_blocks[idx]
-            if (ended):
-                mark_conv3d_ended(up_block)
-            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
-                sample = checkpoint_fn(up_block)(
-                    sample, causal=self.causal, timestep=scaled_timestep
-                )
-            else:
-                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
-
-            if sample is None or sample.shape[2] == 0:
-                return
-
-            total_bytes = sample.numel() * sample.element_size()
-            num_chunks = (total_bytes + max_chunk_size - 1) // max_chunk_size
-
-            if num_chunks == 1:
-                # when we are not chunking, detach our x so the callee can free it as soon as they are done
-                next_sample_ref = [sample]
-                del sample
-                run_up(idx + 1, next_sample_ref, ended)
-                return
-            else:
-                samples = torch.chunk(sample, chunks=num_chunks, dim=2)
-
-                for chunk_idx, sample1 in enumerate(samples):
-                    run_up(idx + 1, [sample1], ended and chunk_idx == len(samples) - 1)
-
-        run_up(0, [sample], True)
+        self.run_up(0, [sample], True, timestep_shift_scale, scaled_timestep, checkpoint_fn, output_buffer, output_offset, max_chunk_size)
 
         return output_buffer
 

comfy/ldm/wan/vae.py

@@ -360,6 +360,43 @@ def __init__(self,
             RMS_norm(out_dim, images=False), nn.SiLU(),
             CausalConv3d(out_dim, output_channels, 3, padding=1))
 
+    def run_up(self, layer_idx, x_ref, feat_cache, feat_idx, out_chunks):
+        x = x_ref[0]
+        x_ref[0] = None
+        if layer_idx >= len(self.upsamples):
+            for layer in self.head:
+                if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                    cache_x = x[:, :, -CACHE_T:, :, :]
+                    x = layer(x, feat_cache[feat_idx[0]])
+                    feat_cache[feat_idx[0]] = cache_x
+                    feat_idx[0] += 1
+                else:
+                    x = layer(x)
+            out_chunks.append(x)
+            return
+
+        layer = self.upsamples[layer_idx]
+        if isinstance(layer, Resample) and layer.mode == 'upsample3d' and x.shape[2] > 1:
+            for frame_idx in range(x.shape[2]):
+                self.run_up(
+                    layer_idx,
+                    [x[:, :, frame_idx:frame_idx + 1, :, :]],
+                    feat_cache,
+                    feat_idx.copy(),
+                    out_chunks,
+                )
+            del x
+            return
+
+        if feat_cache is not None:
+            x = layer(x, feat_cache, feat_idx)
+        else:
+            x = layer(x)
+
+        next_x_ref = [x]
+        del x
+        self.run_up(layer_idx + 1, next_x_ref, feat_cache, feat_idx, out_chunks)
+
     def forward(self, x, feat_cache=None, feat_idx=[0]):
         ## conv1
         if feat_cache is not None:
@@ -380,42 +417,7 @@ def forward(self, x, feat_cache=None, feat_idx=[0]):
 
         out_chunks = []
 
-        def run_up(layer_idx, x_ref, feat_idx):
-            x = x_ref[0]
-            x_ref[0] = None
-            if layer_idx >= len(self.upsamples):
-                for layer in self.head:
-                    if isinstance(layer, CausalConv3d) and feat_cache is not None:
-                        cache_x = x[:, :, -CACHE_T:, :, :]
-                        x = layer(x, feat_cache[feat_idx[0]])
-                        feat_cache[feat_idx[0]] = cache_x
-                        feat_idx[0] += 1
-                    else:
-                        x = layer(x)
-                out_chunks.append(x)
-                return
-
-            layer = self.upsamples[layer_idx]
-            if isinstance(layer, Resample) and layer.mode == 'upsample3d' and x.shape[2] > 1:
-                for frame_idx in range(x.shape[2]):
-                    run_up(
-                        layer_idx,
-                        [x[:, :, frame_idx:frame_idx + 1, :, :]],
-                        feat_idx.copy(),
-                    )
-                del x
-                return
-
-            if feat_cache is not None:
-                x = layer(x, feat_cache, feat_idx)
-            else:
-                x = layer(x)
-
-            next_x_ref = [x]
-            del x
-            run_up(layer_idx + 1, next_x_ref, feat_idx)
-
-        run_up(0, [x], feat_idx)
+        self.run_up(0, [x], feat_cache, feat_idx, out_chunks)
         return out_chunks
 
 

comfy/model_base.py

@@ -285,6 +285,12 @@ def concat_cond(self, **kwargs):
             return data
         return None
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        """Override in subclasses to handle model-specific cond slicing for context windows.
+        Return a sliced cond object, or None to fall through to default handling.
+        Use comfy.context_windows.slice_cond() for common cases."""
+        return None
+
     def extra_conds(self, **kwargs):
         out = {}
         concat_cond = self.concat_cond(**kwargs)
@@ -1375,6 +1381,12 @@ def extra_conds(self, **kwargs):
         out['vace_strength'] = comfy.conds.CONDConstant(vace_strength)
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "vace_context":
+            import comfy.context_windows
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=3, retain_index_list=retain_index_list)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN21_Camera(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.CameraWanModel)
@@ -1427,6 +1439,12 @@ def extra_conds(self, **kwargs):
 
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "audio_embed":
+            import comfy.context_windows
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=1)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN22_Animate(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model_animate.AnimateWanModel)
@@ -1444,6 +1462,14 @@ def extra_conds(self, **kwargs):
             out['pose_latents'] = comfy.conds.CONDRegular(self.process_latent_in(pose_latents))
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        import comfy.context_windows
+        if cond_key == "face_pixel_values":
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_scale=4, temporal_offset=1)
+        if cond_key == "pose_latents":
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_offset=1)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN22_S2V(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)
@@ -1480,6 +1506,12 @@ def extra_conds_shapes(self, **kwargs):
             out['reference_motion'] = reference_motion.shape
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "audio_embed":
+            import comfy.context_windows
+            return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=1)
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 class WAN22(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)

comfy/sd.py

@@ -978,6 +978,7 @@ def decode(self, samples_in, vae_options={}):
             do_tile = True
 
         if do_tile:
+            comfy.model_management.soft_empty_cache()
             dims = samples_in.ndim - 2
             if dims == 1 or self.extra_1d_channel is not None:
                 pixel_samples = self.decode_tiled_1d(samples_in)
@@ -1059,6 +1060,7 @@ def encode(self, pixel_samples):
             do_tile = True
 
         if do_tile:
+            comfy.model_management.soft_empty_cache()
             if self.latent_dim == 3:
                 tile = 256
                 overlap = tile // 4