AV latent support for LTXVLoopingSampler and LTXVExtendSampler

.gitignore

@@ -160,3 +160,11 @@ cython_debug/
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 #.idea/
+
+# macOS
+.DS_Store
+._*
+
+# AI assistant context (local only)
+/CLAUDE.md
+/AGENTS.md

__init__.py

@@ -58,7 +58,7 @@
 from .tiled_vae_decode import LTXVTiledVAEDecode
 from .tricks import NODE_CLASS_MAPPINGS as TRICKS_NODE_CLASS_MAPPINGS
 from .tricks import NODE_DISPLAY_NAME_MAPPINGS as TRICKS_NODE_DISPLAY_NAME_MAPPINGS
-from .utiltily_nodes import FloatToInt, ImageToCPU
+from .utiltily_nodes import FloatToInt, ImageToCPU, LTXVLoopingReferenceSchedule
 from .vae_patcher import LTXVPatcherVAE
 from .vanish_nodes import LTXVDilateVideoMask, LTXVInpaintPreprocess
 
@@ -96,6 +96,7 @@
     "LTXVMultiPromptProvider": MultiPromptProvider,
     "ImageToCPU": ImageToCPU,
     "LTXFloatToInt": FloatToInt,
+    "LTXVLoopingReferenceSchedule": LTXVLoopingReferenceSchedule,
     "LTXVStatNormLatent": LTXVStatNormLatent,
     "LTXVPerStepStatNormPatcher": LTXVPerStepStatNormPatcher,
     "LTXVGemmaCLIPModelLoader": LTXVGemmaCLIPModelLoader,

easy_samplers.py

@@ -16,6 +16,50 @@
 from .nodes_registry import comfy_node
 
 
+def _make_av_latent_dict(video_latent_dict, audio_tensor, audio_noise_mask=None):
+    """Wrap video latent dict + audio tensor into AV latent dict with NestedTensor.
+
+    If audio_tensor is None, returns video_latent_dict unchanged.
+    Creates matching noise masks for both modalities when either is present.
+    """
+    if audio_tensor is None:
+        return video_latent_dict
+    result = video_latent_dict.copy()
+    result["samples"] = NestedTensor([result["samples"], audio_tensor])
+    video_mask = result.get("noise_mask")
+    if video_mask is not None or audio_noise_mask is not None:
+        if video_mask is None:
+            vs = result["samples"].tensors[0]
+            video_mask = torch.ones(
+                vs.shape[0], 1, vs.shape[2], vs.shape[3], vs.shape[4],
+                device=vs.device, dtype=vs.dtype,
+            )
+        if audio_noise_mask is None:
+            audio_noise_mask = torch.ones(
+                audio_tensor.shape[0], 1, audio_tensor.shape[2], audio_tensor.shape[3],
+                device=audio_tensor.device, dtype=audio_tensor.dtype,
+            )
+        result["noise_mask"] = NestedTensor([video_mask, audio_noise_mask])
+    return result
+
+
+def _split_av_latent_dict(latent_dict):
+    """Split AV latent dict into (video_latent_dict, audio_tensor).
+
+    If the latent is not an AV NestedTensor, returns (latent_dict, None).
+    """
+    samples = latent_dict["samples"]
+    if not isinstance(samples, NestedTensor) or len(samples.tensors) < 2:
+        return latent_dict, None
+    result = latent_dict.copy()
+    result["samples"] = samples.tensors[0]
+    audio = samples.tensors[1]
+    nm = result.get("noise_mask")
+    if nm is not None and isinstance(nm, NestedTensor):
+        result["noise_mask"] = nm.tensors[0]
+    return result, audio
+
+
 def _get_raw_conds_from_guider(guider):
     if not hasattr(guider, "raw_conds"):
         if "negative" not in guider.original_conds:
@@ -148,6 +192,7 @@ def sample(
         optional_initialization_latents=None,
         guiding_start_step=0,
         guiding_end_step=1000,
+        _audio_tile=None,
     ):
         guider = copy.copy(guider)
         guider.original_conds = copy.deepcopy(guider.original_conds)
@@ -262,13 +307,15 @@ def sample(
 
         # Denoise the latent video
         print("Denoising with conditioning on sigmas: ", middle_sigmas)
+        _av = _make_av_latent_dict(latents, _audio_tile)
         (output_latents, denoised_output_latents) = SamplerCustomAdvanced().sample(
             noise=noise,
             guider=guider,
             sampler=sampler,
             sigmas=middle_sigmas,
-            latent_image=latents,
+            latent_image=_av,
         )
+        denoised_output_latents, _audio_tile = _split_av_latent_dict(denoised_output_latents)
 
         # Clean up guides if image conditioning was used
         positive, negative, denoised_output_latents = LTXVCropGuides.execute(
@@ -284,13 +331,18 @@ def sample(
                 "Denoising with no conditioning but with classical i2v noise mask on sigmas: ",
                 low_sigmas,
             )
+            _av = _make_av_latent_dict(denoised_output_latents, _audio_tile)
             (_, denoised_output_latents) = SamplerCustomAdvanced().sample(
                 noise=noise,
                 guider=guider,
                 sampler=sampler,
                 sigmas=low_sigmas,
-                latent_image=denoised_output_latents,
+                latent_image=_av,
             )
+            denoised_output_latents, _audio_tile = _split_av_latent_dict(denoised_output_latents)
+
+        if _audio_tile is not None:
+            denoised_output_latents["_audio"] = _audio_tile
 
         return (denoised_output_latents, positive, negative)
 
@@ -399,6 +451,8 @@ def sample(
         guiding_start_step=0,
         guiding_end_step=1000,
         normalize_per_frame=False,
+        _audio_tile=None,
+        _audio_new_init=None,
     ):
         guider = copy.copy(guider)
         guider.original_conds = copy.deepcopy(guider.original_conds)
@@ -412,7 +466,20 @@ def sample(
 
         positive, negative = _get_raw_conds_from_guider(guider)
 
+        # Handle AV latents (standalone mode)
+        _standalone_av = False
+        _accumulated_audio = _audio_tile
         samples = latents["samples"]
+        if isinstance(samples, NestedTensor) and len(samples.tensors) == 2:
+            if _accumulated_audio is None:
+                _accumulated_audio = samples.tensors[1]
+            _standalone_av = True
+            latents = latents.copy()
+            latents["samples"] = samples.tensors[0]
+            if "noise_mask" in latents and isinstance(latents["noise_mask"], NestedTensor):
+                latents["noise_mask"] = latents["noise_mask"].tensors[0]
+            samples = latents["samples"]
+
         batch, channels, frames, height, width = samples.shape
         time_scale_factor, width_scale_factor, height_scale_factor = (
             vae.downscale_index_formula
@@ -428,6 +495,52 @@ def sample(
             latents, -overlap, -1
         )
 
+        # Set up audio extend tile if audio is available
+        _audio_extend_tile = None
+        _audio_noise_mask = None
+        _audio_overlap = 0
+        if _accumulated_audio is not None:
+            audio_T = _accumulated_audio.shape[2]
+            video_T = frames
+            audio_ratio = audio_T / max(video_T, 1)
+            _audio_overlap = max(1, round(overlap * audio_ratio))
+            video_new_latent_frames = num_new_frames // time_scale_factor
+            audio_new_frames = max(1, round(video_new_latent_frames * audio_ratio))
+
+            # Build audio tile: overlap (already denoised) + new frames.
+            # If _audio_new_init is provided (stage-2 refinement), use it
+            # as initialization for the new frames instead of zeros.
+            audio_overlap_data = _accumulated_audio[:, :, -_audio_overlap:]
+            if _audio_new_init is not None:
+                available = min(audio_new_frames, _audio_new_init.shape[2])
+                audio_new_data = _audio_new_init[:, :, :available].clone()
+                if available < audio_new_frames:
+                    pad = torch.zeros(
+                        _accumulated_audio.shape[0], _accumulated_audio.shape[1],
+                        audio_new_frames - available, _accumulated_audio.shape[3],
+                        device=_accumulated_audio.device, dtype=_accumulated_audio.dtype,
+                    )
+                    audio_new_data = torch.cat([audio_new_data, pad], dim=2)
+            else:
+                audio_new_data = torch.zeros(
+                    _accumulated_audio.shape[0], _accumulated_audio.shape[1],
+                    audio_new_frames, _accumulated_audio.shape[3],
+                    device=_accumulated_audio.device, dtype=_accumulated_audio.dtype,
+                )
+            _audio_extend_tile = torch.cat([audio_overlap_data, audio_new_data], dim=2)
+
+            # Audio noise mask: preserve overlap, denoise new
+            _audio_noise_mask = torch.ones(
+                _audio_extend_tile.shape[0], 1,
+                _audio_extend_tile.shape[2], _audio_extend_tile.shape[3],
+                device=_audio_extend_tile.device, dtype=_audio_extend_tile.dtype,
+            )
+            _audio_noise_mask[:, :, :_audio_overlap] = 1.0 - strength
+            print(
+                f"[ExtendSampler] Audio extend tile: overlap={_audio_overlap}, "
+                f"new={audio_new_frames}, total={_audio_extend_tile.shape[2]}"
+            )
+
         if optional_initialization_latents is None:
             new_latents = EmptyLTXVLatentVideo.execute(
                 width=width * width_scale_factor,
@@ -488,13 +601,15 @@ def sample(
         if len(high_sigmas) > 1:
             guider.set_conds(positive, negative)
             print("Denoising with overlap conditioning only on sigmas: ", high_sigmas)
+            _av = _make_av_latent_dict(new_latents, _audio_extend_tile, _audio_noise_mask)
             (_, new_latents) = SamplerCustomAdvanced().sample(
                 noise=noise,
                 guider=guider,
                 sampler=sampler,
                 sigmas=high_sigmas,
-                latent_image=new_latents,
+                latent_image=_av,
             )
+            new_latents, _audio_extend_tile = _split_av_latent_dict(new_latents)
 
         if optional_guiding_latents is not None:
             optional_guiding_latents = LTXVSelectLatents().select_latents(
@@ -533,13 +648,15 @@ def sample(
 
         # Denoise the latent video
         print("Denoising with full conditioning on sigmas: ", middle_sigmas)
+        _av = _make_av_latent_dict(new_latents, _audio_extend_tile, _audio_noise_mask)
         (output_latents, denoised_output_latents) = SamplerCustomAdvanced().sample(
             noise=noise,
             guider=guider,
             sampler=sampler,
             sigmas=middle_sigmas,
-            latent_image=new_latents,
+            latent_image=_av,
         )
+        denoised_output_latents, _audio_extend_tile = _split_av_latent_dict(denoised_output_latents)
 
         positive, negative, denoised_output_latents = LTXVCropGuides.execute(
             positive=positive,
@@ -591,13 +708,15 @@ def sample(
                 "Denoising with overlap + keyframes conditioning only on sigmas: ",
                 low_sigmas,
             )
+            _av = _make_av_latent_dict(denoised_output_latents, _audio_extend_tile, _audio_noise_mask)
             (_, denoised_output_latents) = SamplerCustomAdvanced().sample(
                 noise=noise,
                 guider=guider,
                 sampler=sampler,
                 sigmas=low_sigmas,
-                latent_image=denoised_output_latents,
+                latent_image=_av,
             )
+            denoised_output_latents, _audio_extend_tile = _split_av_latent_dict(denoised_output_latents)
             positive, negative, denoised_output_latents = LTXVCropGuides.execute(
                 positive=positive,
                 negative=negative,
@@ -621,6 +740,16 @@ def sample(
         (latents,) = LinearOverlapLatentTransition().process(
             latents, truncated_denoised_output_latents, overlap - 1, axis=2
         )
+
+        # Accumulate audio: append new (non-overlap) audio frames
+        if _accumulated_audio is not None and _audio_extend_tile is not None:
+            new_audio = _audio_extend_tile[:, :, _audio_overlap:]
+            accumulated_audio_out = torch.cat([_accumulated_audio, new_audio], dim=2)
+            if _standalone_av:
+                latents["samples"] = NestedTensor([latents["samples"], accumulated_audio_out])
+            else:
+                latents["_audio"] = accumulated_audio_out
+
         return (latents, positive, negative)
 
 
@@ -692,6 +821,7 @@ def sample(
         guiding_strength=1.0,
         guiding_start_step=0,
         guiding_end_step=1000,
+        _audio_tile=None,
     ):
         guider = copy.copy(guider)
         guider.original_conds = copy.deepcopy(guider.original_conds)
@@ -735,13 +865,15 @@ def sample(
                 "Denoising with keyframes only [if available] on sigmas: ",
                 high_sigmas,
             )
+            _av = _make_av_latent_dict(new_latents, _audio_tile)
             (_, new_latents) = SamplerCustomAdvanced().sample(
                 noise=noise,
                 guider=guider,
                 sampler=sampler,
                 sigmas=high_sigmas,
-                latent_image=new_latents,
+                latent_image=_av,
             )
+            new_latents, _audio_tile = _split_av_latent_dict(new_latents)
 
         if optional_cond_indices is not None and 0 in optional_cond_indices:
             guiding_latents = LTXVSelectLatents().select_latents(
@@ -806,13 +938,15 @@ def sample(
 
         # Denoise the latent video
         print("Denoising with full conditioning on sigmas: ", middle_sigmas)
+        _av = _make_av_latent_dict(new_latents, _audio_tile)
         (_, denoised_output_latents) = SamplerCustomAdvanced().sample(
             noise=noise,
             guider=guider,
             sampler=sampler,
             sigmas=middle_sigmas,
-            latent_image=new_latents,
+            latent_image=_av,
         )
+        denoised_output_latents, _audio_tile = _split_av_latent_dict(denoised_output_latents)
 
         # Clean up guides if image conditioning was used
         positive, negative, denoised_output_latents = LTXVCropGuides.execute(
@@ -827,19 +961,24 @@ def sample(
                 "Denoising with keyframes only [if available] conditioning on sigmas: ",
                 low_sigmas,
             )
+            _av = _make_av_latent_dict(denoised_output_latents, _audio_tile)
             (_, denoised_output_latents) = SamplerCustomAdvanced().sample(
                 noise=noise,
                 guider=guider,
                 sampler=sampler,
                 sigmas=low_sigmas,
-                latent_image=denoised_output_latents,
+                latent_image=_av,
             )
+            denoised_output_latents, _audio_tile = _split_av_latent_dict(denoised_output_latents)
             positive, negative, denoised_output_latents = LTXVCropGuides.execute(
                 positive=positive,
                 negative=negative,
                 latent=denoised_output_latents,
             )
 
+        if _audio_tile is not None:
+            denoised_output_latents["_audio"] = _audio_tile
+
         return (denoised_output_latents, positive, negative)