Skip to content

LTX-Video Support for ComfyUI

Automated PR - 2026-01-15

sync-bot683d46186a13
diff --git a/__init__.py b/__init__.py
index dfef2de..33c1d53 100644
--- a/__init__.py
+++ b/__init__.py
@@ -5,6 +5,7 @@ from .easy_samplers import (
LTXVBaseSampler,
LTXVExtendSampler,
LTXVInContextSampler,
+ LTXVNormalizingSampler,
)
from .gemma_encoder import LTXVGemmaCLIPModelLoader, LTXVGemmaEnhancePrompt
from .guide import LTXVAddGuideAdvanced
@@ -61,6 +62,7 @@ NODE_CLASS_MAPPINGS = {
"LTXVBaseSampler": LTXVBaseSampler,
"LTXVInContextSampler": LTXVInContextSampler,
"LTXVExtendSampler": LTXVExtendSampler,
+ "LTXVNormalizingSampler": LTXVNormalizingSampler,
"LTXVPreprocessMasks": LTXVPreprocessMasks,
"LTXVPatcherVAE": LTXVPatcherVAE,
"LTXVPromptEnhancer": LTXVPromptEnhancer,
diff --git a/easy_samplers.py b/easy_samplers.py
index 84d7fd0..d44e77c 100644
--- a/easy_samplers.py
+++ b/easy_samplers.py
@@ -4,6 +4,9 @@ import comfy
import comfy_extras
import nodes
import torch
+from comfy.ldm.lightricks.av_model import LTXAVModel
+from comfy.nested_tensor import NestedTensor
+from comfy_api.latest import io
from comfy_extras.nodes_custom_sampler import SamplerCustomAdvanced, SplitSigmas
from comfy_extras.nodes_lt import EmptyLTXVLatentVideo, LTXVAddGuide, LTXVCropGuides
@@ -907,3 +910,142 @@ class LinearOverlapLatentTransition:
"batch_index": combined_batch_index,
},
)
+
+
+@comfy_node(description="LTXV Normalizing Sampler")
+class LTXVNormalizingSampler(io.ComfyNode):
+ @classmethod
+ def define_schema(cls):
+ return io.Schema(
+ node_id="LTXVNormalizingSampler",
+ category="utility",
+ inputs=[
+ io.Noise.Input("noise"),
+ io.Guider.Input("guider"),
+ io.Sampler.Input("sampler"),
+ io.Sigmas.Input("sigmas"),
+ io.Latent.Input("latent_image"),
+ io.String.Input(
+ "video_normalization_factors", default="1,1,1,1,1,1,1,1"
+ ),
+ io.String.Input(
+ "audio_normalization_factors", default="1,1,0.25,1,1,0.25,1,1"
+ ),
+ ],
+ outputs=[
+ io.Latent.Output(display_name="denoised_output"),
+ ],
+ )
+
+ @classmethod
+ def execute(
+ cls,
+ noise,
+ guider,
+ sampler,
+ sigmas,
+ latent_image,
+ video_normalization_factors,
+ audio_normalization_factors,
+ ) -> io.NodeOutput:
+
+ if (
+ guider.model_patcher.model.diffusion_model.__class__.__name__
+ != "LTXAVModel"
+ ):
+ raise ValueError()
+
+ ltxav: LTXAVModel = guider.model_patcher.model.diffusion_model
+
+ video_normalization_factors = video_normalization_factors.split(",")
+ audio_normalization_factors = audio_normalization_factors.split(",")
+ video_normalization_factors = [
+ float(factor) for factor in video_normalization_factors
+ ]
+ audio_normalization_factors = [
+ float(factor) for factor in audio_normalization_factors
+ ]
+
+ # Extend normalization factors to match the length of sigmas
+ sigmas_len = len(sigmas) - 1
+ if len(video_normalization_factors) < sigmas_len:
+ if len(video_normalization_factors) > 0:
+ video_normalization_factors.extend(
+ [video_normalization_factors[-1]]
+ * (sigmas_len - len(video_normalization_factors))
+ )
+ if len(audio_normalization_factors) < sigmas_len:
+ if len(audio_normalization_factors) > 0:
+ audio_normalization_factors.extend(
+ [audio_normalization_factors[-1]]
+ * (sigmas_len - len(audio_normalization_factors))
+ )
+
+ # Calculate indices where both normalization factors are not 1.0
+ sampling_split_indices = [
+ i + 1
+ for i, (v, a) in enumerate(
+ zip(video_normalization_factors, audio_normalization_factors)
+ )
+ if v != 1.0 or a != 1.0
+ ]
+ print("Sampling split indices: %s" % sampling_split_indices, flush=True)
+
+ # Split sigmas according to sampling_split_indices
+ def split_by_indices(arr, indices):
+ """
+ Splits arr into chunks according to indices (split points).
+ Indices are treated as starting a new chunk at each index in the list.
+ """
+ if not indices:
+ return [arr]
+ split_points = sorted(set(indices))
+ chunks = []
+ prev = 0
+ for idx in split_points:
+ if prev < idx:
+ chunks.append(arr[prev : idx + 1])
+ prev = idx
+ if prev < len(arr):
+ chunks.append(arr[prev:])
+ return chunks
+
+ sigmas_chunks = split_by_indices(sigmas, sampling_split_indices)
+ print("Sigmas chunks: %s" % sigmas_chunks, flush=True)
+
+ i = 0
+ for sigmas_chunk in sigmas_chunks:
+ i += len(sigmas_chunk) - 1
+ print("Sampling with sigmas %s" % (sigmas_chunk), flush=True)
+ (_, latent_image) = SamplerCustomAdvanced().execute(
+ noise=noise,
+ guider=guider,
+ sampler=sampler,
+ sigmas=sigmas_chunk,
+ latent_image=latent_image,
+ )
+ video_samples, audio_samples = ltxav.separate_audio_and_video_latents(
+ latent_image["samples"].tensors,
+ None,
+ )
+ if i - 1 < len(video_normalization_factors) and i - 1 < len(
+ audio_normalization_factors
+ ):
+ video_samples = video_samples * video_normalization_factors[i - 1]
+ audio_samples = audio_samples * audio_normalization_factors[i - 1]
+ latent_image["samples"] = NestedTensor(
+ ltxav.recombine_audio_and_video_latents(
+ video_samples, audio_samples
+ )
+ )
+ print(
+ "After %d steps, the latent image was normalized by %f and %f"
+ % (
+ i,
+ video_normalization_factors[i - 1],
+ audio_normalization_factors[i - 1],
+ ),
+ flush=True,
+ )
+
+ return io.NodeOutput(latent_image)
diff --git a/example_workflows/LTX-2_I2V_Distilled_wLora.json b/example_workflows/LTX-2_I2V_Distilled_wLora.json
index 6075a08..6bb9864 100644
--- a/example_workflows/LTX-2_I2V_Distilled_wLora.json
+++ b/example_workflows/LTX-2_I2V_Distilled_wLora.json
@@ -565,7 +565,7 @@
"title": "Video Size",
"properties": {},
"widgets_values": [
- "Width & height settings must be divisible by 32 + 1. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
+ "Width & height settings must be divisible by 64. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
],
"color": "#432",
"bgcolor": "#653"
diff --git a/example_workflows/LTX-2_I2V_Full_wLora.json b/example_workflows/LTX-2_I2V_Full_wLora.json
index 63b9611..4e71128 100644
--- a/example_workflows/LTX-2_I2V_Full_wLora.json
+++ b/example_workflows/LTX-2_I2V_Full_wLora.json
@@ -743,7 +743,7 @@
"title": "Video Size",
"properties": {},
"widgets_values": [
- "Width & height settings must be divisible by 32 + 1. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
+ "Width & height settings must be divisible by 64. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
],
"color": "#432",
"bgcolor": "#653"
diff --git a/example_workflows/LTX-2_ICLoRA_All_Distilled.json b/example_workflows/LTX-2_ICLoRA_All_Distilled.json
index 2395d9d..faf2ab9 100644
--- a/example_workflows/LTX-2_ICLoRA_All_Distilled.json
+++ b/example_workflows/LTX-2_ICLoRA_All_Distilled.json
@@ -57,7 +57,7 @@
"title": "Video Size",
"properties": {},
"widgets_values": [
- "Width & height settings must be divisible by 32 + 1. \n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
+ "Width & height settings must be divisible by 64. \n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
],
"color": "#432",
"bgcolor": "#653"
diff --git a/example_workflows/LTX-2_T2V_Distilled_wLora.json b/example_workflows/LTX-2_T2V_Distilled_wLora.json
index d0a1f00..b0a524d 100644
--- a/example_workflows/LTX-2_T2V_Distilled_wLora.json
+++ b/example_workflows/LTX-2_T2V_Distilled_wLora.json
@@ -567,7 +567,7 @@
"title": "Video Size",
"properties": {},
"widgets_values": [
- "Width & height settings must be divisible by 32 + 1. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
+ "Width & height settings must be divisible by 64. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
],
"color": "#432",
"bgcolor": "#653"
diff --git a/example_workflows/LTX-2_T2V_Full_wLora.json b/example_workflows/LTX-2_T2V_Full_wLora.json
index 5a7f6a6..5f0fd32 100644
--- a/example_workflows/LTX-2_T2V_Full_wLora.json
+++ b/example_workflows/LTX-2_T2V_Full_wLora.json
@@ -520,7 +520,7 @@
"title": "Video Size",
"properties": {},
"widgets_values": [
- "Width & height settings must be divisible by 32 + 1. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
+ "Width & height settings must be divisible by 64. \nFrame count must be divisible by 8 + 1. \n\n\nRunning with invalid parameters **will not cause errors**. Instead, the flow will silently choose the closest valid parameters. "
],
"color": "#432",
"bgcolor": "#653"
diff --git a/gemma_encoder.py b/gemma_encoder.py
index 43031c6..3524ed7 100644
--- a/gemma_encoder.py
+++ b/gemma_encoder.py
@@ -1,5 +1,6 @@
import json
import logging
+from glob import glob
from pathlib import Path
from typing import Optional
@@ -528,8 +529,11 @@ def find_matching_dir(root_path: str, pattern: str) -> str:
"""
Recursively search for files matching a glob pattern and return the parent directory of the first match.
"""
-
- matches = list(Path(root_path).rglob(pattern))
+ matches = [
+ Path(p)
+ for p in glob(f"{root_path}/**", recursive=True)
+ if Path(p).match(pattern)
+ ]
if not matches:
raise FileNotFoundError(
f"No files matching pattern '{pattern}' found under {root_path}"
diff --git a/guiders/multimodal_guider.py b/guiders/multimodal_guider.py
index fa2b604..441263b 100644
--- a/guiders/multimodal_guider.py
+++ b/guiders/multimodal_guider.py
@@ -85,6 +85,20 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
return list(stg_indexes)
+ def unpack_latents(self, x: torch.Tensor):
+ latent_shapes = (
+ self.conds.get("positive", {})[0]
+ .get("model_conds", {})
+ .get("latent_shapes", None)
+ .cond
+ )
+ vx, ax = comfy.utils.unpack_latents(x, latent_shapes)
+ return vx, ax
+
+ def pack_latents(self, vx: torch.Tensor, ax: torch.Tensor):
+ x, latent_shapes = comfy.utils.pack_latents([vx, ax])
+ return x, latent_shapes
+
def predict_noise(
self,
x: torch.Tensor,
@@ -92,7 +106,6 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
model_options: dict = {},
seed=None,
):
- x_shape = x.shape
# in CFGGuider.predict_noise, we call sampling_function(), which uses cfg_function() to compute pos & neg
# but we'd rather do a single batch of sampling pos, neg, and perturbed, so we call calc_cond_batch([perturbed,pos,neg]) directly
current_step = self.current_step
@@ -115,16 +128,13 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
run_a2v = video_params.do_cross_attn(current_step)
run_v2a = audio_params.do_cross_attn(current_step)
- vx, ax = self.inner_model.diffusion_model.separate_audio_and_video_latents(
- x, audio_length
- )
+ vx, ax = self.unpack_latents(x)
if not run_vx:
vx = self.last_denoised_v
if not run_ax:
ax = self.last_denoised_a
- x = self.inner_model.diffusion_model.recombine_audio_and_video_latents(
- vx, ax, x_shape
- )
+ x, _ = self.pack_latents(vx, ax)
+
if not run_vx and not run_ax:
return x
@@ -145,11 +155,7 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
del model_options["transformer_options"]["run_ax"]
del model_options["transformer_options"]["a2v_cross_attn"]
del model_options["transformer_options"]["v2a_cross_attn"]
- v_noise_pred_pos, a_noise_pred_pos = (
- self.inner_model.diffusion_model.separate_audio_and_video_latents(
- noise_pred_pos, audio_length
- )
- )
+ v_noise_pred_pos, a_noise_pred_pos = self.unpack_latents(noise_pred_pos)
a_noise_pred_neg, v_noise_pred_neg = 0, 0
a_noise_pred_perturbed, v_noise_pred_perturbed = 0, 0
@@ -168,11 +174,7 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
timestep,
model_options,
)[0]
- v_noise_pred_neg, a_noise_pred_neg = (
- self.inner_model.diffusion_model.separate_audio_and_video_latents(
- noise_pred_neg, audio_length
- )
- )
+ v_noise_pred_neg, a_noise_pred_neg = self.unpack_latents(noise_pred_neg)
finally:
del model_options["transformer_options"]["run_vx"]
del model_options["transformer_options"]["run_ax"]
@@ -202,10 +204,8 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
timestep,
model_options,
)[0]
- v_noise_pred_perturbed, a_noise_pred_perturbed = (
- self.inner_model.diffusion_model.separate_audio_and_video_latents(
- noise_pred_perturbed, audio_length
- )
+ v_noise_pred_perturbed, a_noise_pred_perturbed = self.unpack_latents(
+ noise_pred_perturbed
)
finally:
self.stg_flag.do_skip = False
@@ -229,10 +229,8 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
timestep,
model_options,
)[0]
- v_noise_pred_modality, a_noise_pred_modality = (
- self.inner_model.diffusion_model.separate_audio_and_video_latents(
- noise_pred_modality, audio_length
- )
+ v_noise_pred_modality, a_noise_pred_modality = self.unpack_latents(
+ noise_pred_modality
)
finally:
del model_options["transformer_options"]["a2v_cross_attn"]
@@ -260,9 +258,7 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
else:
ax = self.last_denoised_a
- x = self.inner_model.diffusion_model.recombine_audio_and_video_latents(
- vx, ax, x_shape
- )
+ x, _ = self.pack_latents(vx, ax)
# normally this would be done in cfg_function, but we skipped
# that for efficiency: we can compute the noise predictions in
@@ -285,11 +281,7 @@ class MultimodalGuider(comfy.samplers.CFGGuider):
}
x = fn(args)
- self.last_denoised_v, self.last_denoised_a = (
- self.inner_model.diffusion_model.separate_audio_and_video_latents(
- x, audio_length
- )
- )
+ self.last_denoised_v, self.last_denoised_a = self.unpack_latents(x)
return x