fix: clone latent before TAESD preview decode to prevent corruption

[octo-patch]

Fixes #13366

Problem

The TAESD preview decoders (TAESDPreviewerImpl and TAEHVPreviewerImpl) were passing tensor views (slices of x0) directly to the decoder's decode() call. Since these are views (not copies), any in-place operation within the decoder's pipeline could potentially write back through the view and corrupt the original x0 latent mid-sampling.

This was specifically observed when:

• Using a model with the Wan21 latent format (e.g., QwenImage)
• lighttaew2_1.safetensors (WanVAE) is present in models/vae_approx/
• TAESD preview method is enabled

In this configuration, the full VAE decode pipeline (including model management, conv operations, and CUDA kernel scheduling) processes the non-contiguous tensor slice, which can result in x0 being modified in-place.

Solution

Clone the tensor slice before passing it to the decoder:

# Before
x_sample = self.taesd.decode(x0[:1, :, :1])[0][0]

# After
x_sample = self.taesd.decode(x0[:1, :, :1].clone())[0][0]

This ensures the original x0 latent is never modified by any operations within the preview decode path, including any potential in-place writes in CUDA kernels or PyTorch operations that may operate directly on the underlying memory of non-contiguous views.

Testing

• The fix is minimal and targeted: only the input to the preview decoder is cloned
• No change to preview output quality or logic
• The clone creates an independent contiguous copy, protecting x0 from any potential modification during the preview decode phase

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 02a3f3bf-5825-4e44-99b6-c51484cda844

📥 Commits

Reviewing files that changed from the base of the PR and between c2657d5 and ce6b8b7.

📒 Files selected for processing (1)

• latent_preview.py

---

📝 Walkthrough

Walkthrough

The pull request modifies the latent preview decoding implementations in latent_preview.py. Specifically, in both TAESDPreviewerImpl.decode_latent_to_preview and TAEHVPreviewerImpl.decode_latent_to_preview methods, latent tensor slices are now cloned before being passed to the decoder. This prevents the decoder from modifying the original latent tensor through in-place operations. The changes are internal implementation details with no alterations to public or exported entity declarations.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately and concisely describes the primary change: cloning latent tensors before TAESD preview decode to prevent corruption.
Description check	✅ Passed	The description is thorough and directly related to the changeset, explaining the problem, solution, and testing approach for the latent tensor corruption fix.
Linked Issues check	✅ Passed	The PR successfully addresses issue #13366 by implementing the required fix: cloning tensor slices before passing to TAESD decoder to prevent in-place modifications of the original latent.
Out of Scope Changes check	✅ Passed	All changes are in-scope and directly address the linked issue; only the specific latent cloning logic is modified in the two previewer implementations.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}