[mirzasnowy]

Fixes #15139.

Bug

With variance_type="fixed_large" and timestep_spacing="trailing", DDPMScheduler.step() adds noise to the final clean-data prediction. For 1000 training timesteps and 50 inference steps, the last selected timestep is 19 and previous_timestep(19) returns -1, but step() still enters its noise branch because it checks t > 0. The mean at this boundary is already pred_original_sample (alpha_prod_t_prev = 1), so the fresh random term corrupts the returned clean prediction — and changing only the random seed changes the returned clean endpoint. The same guard exists in DDPMParallelScheduler.step().

Fix

In both step() methods, the noise branch guard changes from t > 0 to prev_t >= 0. prev_t is already computed at the top of each method and already used the same way for the alpha coefficients (alpha_prod_t_prev = ... if prev_t >= 0 else self.one), so the new condition is consistent with the method's own terminal handling. Behavior changes only at terminal steps with t > 0 (trailing spacing or custom schedules ending nonzero); all other steps are bit-identical. The parallel scheduler's step() is not a # Copied from block (verified against the file's copied-block list), so no make fix-copies sync is owed; both guards were changed identically regardless.

Tests

• Issue repro (CPU, public scheduler API, no model weights): before fix, fixed_large 19 -1 deviates 0.1654895395040512 for both schedulers (fixed_small 0.0); after fix, all four combinations 0.0.
• New test_terminal_step_adds_no_noise in both test files: FAIL on pristine code, PASS with the fix.
• Neighbouring suites test_scheduler_ddpm.py + test_scheduler_ddpm_parallel.py: 77 passed. ruff check clean with the repo's own pyproject.toml; ruff format --check clean on the touched files.

Before submitting

• Did you use an AI agent 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?

Self-review notes (final round)

Rubric: .ai/references/review-rules.md (plus code_style copied-code rule).

Blocking issues

None.

Non-blocking issues

1. Maintainer-preferred formulation (leave for the actual review).
   The issue itself says "This needs maintainer review before proposing a PR."
   This fix follows the reporter's suggested approach: gate noise addition on predecessor existence (prev_t >= 0) instead of on t > 0. That condition is already how the same methods handle the terminal case two lines above (alpha_prod_t_prev = ... if prev_t >= 0 else self.one), so it is the most consistent formulation available. If a maintainer prefers a different formulation (e.g. comparing against the last timestep), the regression tests in this diff still define the required behavior exactly.
2. Variance-mode coverage (not a gap, noted for completeness).
   The guard change sits above all variance branches (fixed_small, fixed_large, fixed_small_log, learned_range), so it applies uniformly; the reporter reproduced the bug with learned/learned_range as well. The regression tests cover fixed_small/fixed_large (the issue's exact repro). A learned_range case would need a doubled-channel model output to reach the split path; it exercises the same shared guard, so it is not added.

Dead code (advisory)

None added. The change narrows when an existing branch runs; it adds no new code paths, parameters, or classes.

Verdict

READY — fix before submitting: nothing. Leave for the actual review: item 1 above (maintainer-preferred formulation).

I have reviewed every line of this diff.