LTXAV: SamplerCustomAdvanced crashes with tensor-shape mismatch when latent is nested but noise is not

[djdarcy]

SamplerCustomAdvanced crashes with RuntimeError: The size of tensor a (N) must match the size of tensor b (M) at non-singleton dimension 2 when sampling LTXAV (audio+video) workflows that use LTXVConcatAVLatent to combine video and audio latents.

Still reproduces on master as of 2806163f (2026-05-03).

Repro

1. Build (or load) any LTXAV workflow that:
   • Combines video and audio latents via LTXVConcatAVLatent, and
   • Feeds the combined latent into SamplerCustomAdvanced whose noise input comes from a generator that produces a non-nested noise tensor (e.g. plain RandomNoise reading the upstream video latent).
2. Queue the prompt.

Example workflow that triggers it: any popular LTX-2.3 I2V/T2V workflow built around LTXVConcatAVLatent + SamplerCustomAdvanced.

Stack trace (abbreviated)

File "comfy_extras/nodes_custom_sampler.py", line 963, in execute
  samples = guider.sample(noise.generate_noise(latent), latent_image, sampler, sigmas, ...)
File "comfy/samplers.py", line 1052, in sample
  output = executor.execute(noise, latent_image, sampler, sigmas, ...)
...
File "comfy/samplers.py", line 744, in sample
  noise = model_wrap.inner_model.model_sampling.noise_scaling(sigmas[0], noise, latent_image, ...)
File "comfy/model_sampling.py", line 96, in noise_scaling
  return sigma * noise + (1.0 - sigma) * latent_image
RuntimeError: The size of tensor a (6820) must match the size of tensor b (905216) at non-singleton dimension 2

Reason

In CFGGuider.sample() at comfy/samplers.py:1008-1010:

if latent_image.is_nested:
    latent_image, latent_shapes = comfy.utils.pack_latents(latent_image.unbind())
    noise, _ = comfy.utils.pack_latents(noise.unbind())

When latent_image is a NestedTensor but noise is a plain tensor, noise.unbind() splits along dim 0 (channels) producing far more tensors than expected. After pack_latents flattens and concatenates, noise and latent_image end up with mismatched dim-2 sizes, and noise_scaling crashes on the broadcast.

The denoise_mask block immediately below already handles this asymmetry defensively (checks is_nested, truncates extras, pads missing components with torch.ones). The noise branch should do the same with torch.zeros_like.

Possible solution?

PR #13318 (open since 2026-04-07) implements this. Just bumping for visibility — the fix is small, safe, and mirrors a pattern already used a few lines below.

Presumably why this hasn't been widely reported

• LTXAV is semi-recent and I'm guessing most LTX workflows are video-only, so they never construct a nested latent?
• Among LTXAV workflows, only those whose noise generator returns a non-nested tensor for a nested latent hit this. Workflows that source noise from the same nested latent should be unaffected.

Related

• PR: Fix nested tensor noise mismatch in CFGGuider.sample #13318
• Affected file: comfy/samplers.py (CFGGuider.sample)

[drozbay]

I can't replicate this locally using your workflow. I am not using the SmartResolutionCalc node or GGUF, but I tried to keep it the same otherwise.

[djdarcy]

@drozbay Hrm, can you share the changes you made to the workflow with the exact model used and which ComfyUI branch/commit hash you used? I'll revert the PR and see if perhaps it's model related (the SmartResCalc node just extracts width/height so it shouldn't have anything to do with the issue).

[ywen35951-sys]

ComfyUI 0.28.3 正常
ComfyUI 0.29.0 崩溃
节点：LTXVLoopingSampler + LTXAV
错误：cannot reshape tensor of 0 elements into shape [1, 0, 32, -1]

[djdarcy]

Found the bug in comfyui_smznodes. It replaces comfy.sample.prepare_noise on import (so it applies even if you never place an smZ node), and its version doesn't handle NestedTensor. Models with paired latent streams get noise for the first stream only, and the sampler fails packing it against the full latent: The size of tensor a (N) must match the size of tensor b (M) at non-singleton dimension 2. Affects LTX-AV / LTX-2, MiniMax-H3 and TripoSplat.

This was already diagnosed back in January in shiimizu/ComfyUI_smZNodes#135 . That thread has a one-line workaround if you need it now. Proper fixes: shiimizu/ComfyUI_smZNodes#139 and shiimizu/ComfyUI_smZNodes#140.

Closing this and #13318. Probably also explains #11653.