Do not watermark or postprocess SDXL latent output

[Devam0311]

Issue link: this addresses Issue 4 of #13610.
I have deliberately not used a Fixes keyword: #13610 tracks 7 separate findings, so a closing keyword would close it as soon as this one merges while the others are still open. Happy for the no-issue-needed label to be applied, or I can add a closing keyword if you would rather the review issue be closed and reopened.

What does this PR do?

Fixes Issue 4 of #13610 (stable_diffusion_xl model/pipeline review).

With output_type="latent" the SDXL img2img pipeline sets image = latents and then runs the watermarker and image_processor.postprocess() over it, so raw latents are handled as if they were decoded RGB.

The text-to-image pipeline already guards both:

# pipeline_stable_diffusion_xl.py
if not output_type == "latent":
    if self.watermark is not None:
        image = self.watermark.apply_watermark(image)
    image = self.image_processor.postprocess(image, output_type=output_type)

With a real watermarker this can corrupt or fail on larger latent tensors; with any custom watermarker it is simply called with the wrong data.

Solution

Applies the text-to-image guard to the sites that lack it.

The issue names two. I swept the codebase for the same unguarded shape and found one more — pipeline_pag_sd_xl_img2img.py — so all three are fixed together:

• pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py
• modular_pipelines/stable_diffusion_xl/decoders.py
• pipelines/pag/pipeline_pag_sd_xl_img2img.py (found by sweep, not listed in the issue)

The SDXL inpaint pipeline was checked and already guards correctly.

Testing

test_latent_output_skips_watermark_and_postprocess installs a watermarker that raises if called, requests output_type="latent", and asserts the result is still a latent tensor — checking the channel count, since postprocess would have collapsed it to 3 channels.

Verified it catches the bug: with the fix reverted it fails.

pytest tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl_img2img.py -k "not ip_adapter"
  -> 45 passed, 28 skipped
pytest .../test_stable_diffusion_xl_img2img.py .../test_stable_diffusion_xl.py -k "not ip_adapter"
  -> 110 passed, 65 skipped
ruff check / ruff format --check   -> clean
git diff --check                   -> clean

The TestStableDiffusionXL*PipelineIPAdapter classes are deselected above: they fail on a clean checkout of main in my environment with a ModuleNotFoundError from an optional dependency I do not have installed, unrelated to this change.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Was this discussed/approved via a GitHub issue? — stable_diffusion_xl model/pipeline review #13610
• Did you write any new necessary tests?

Self-review notes

Reviewed against .ai/references/review-rules.md and pipelines.md. No blocking issues.

For the reviewer:

1. This changes what output_type="latent" returns in these pipelines — previously postprocessed, now raw latents. That is the documented and intended behaviour and matches text-to-image, but it is a behaviour change for anyone who was relying on the postprocessed result.
2. I included one file beyond the issue's list (pipeline_pag_sd_xl_img2img.py), found by sweeping for the identical unguarded pattern. Happy to drop it if you'd rather keep this to the two files named.
3. The modular decoder uses hasattr(components, "watermark"); I left that guard as-is and only wrapped it in the output_type check.
4. Scope kept to Issue 4. Issue 2 of the same review is in Use negative crop coordinates when building SDXL negative time ids #14676.

Who can review?

@yiyixuxu @asomoza @hlky

[github-actions]

Hi @Devam0311, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder.

Please note that PRs without a linked issue are likely to be automatically closed 10 days after this notice.

Once the PR links an issue (or gets the no-issue-needed label), you can ignore this message — it stays here as a comment, but it no longer applies.

[adi-IL]

Tested locally on Linux x86_64.

Before this change, running output_type='latent' on StableDiffusionXLImg2ImgPipeline, PAG img2img, and the modular SDXL decoder step passed raw 4-channel latents through RGB watermarking and image_processor.postprocess().

Guarding both behind output_type != 'latent' matches the behavior in the base text-to-image SDXL pipeline and preserves the latent dimensions. The new test asserting watermark bypass passes.