Derive DDIM prev_timestep from the inference schedule

[MohammadSadeghSalehi]

Summary

DDIMScheduler.step formed the previous diffusion index with a uniform stride:

prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps

That expression assumes every inference step advances by the same integer. set_timesteps with timestep_spacing="linspace" does not produce that schedule. For 1000 train steps and 10 inference steps the materialised list strides by 111 while the expression strides by 100, so every step reads alpha_prod_t_prev from the wrong index.

DDIMScheduler defaults to timestep_spacing="leading". Only callers that explicitly select linspace change behaviour. Default sampling outputs do not move.

Proof that leading and trailing are no-ops

Full deterministic loops (eta=0) comparing schedule lookup against the historical stride, same residuals:

spacing	bit-identical	max abs diff
leading	True	0.000000e+00
trailing	True	0.000000e+00
linspace	False	2.584383e+02

On leading and trailing, intermediate prev values match the old stride exactly. At the final step the formula can yield a large negative (for example -100) while the list yields -1; both are < 0 and both select final_alpha_cumprod, so the sample is still bit-identical. On linspace, 9 of 10 steps change alpha_prod_t_prev. Example prev sequence for 1000/10 linspace: schedule [888, 777, …, 0, -1] versus hardcoded [899, 788, …, 11, -100].

Change

Add previous_timestep() that returns the next entry of self.timesteps (last entry → -1). Use it from step (and the parallel batch / variance paths). Same helper is # Copied from into the three DDIM-family schedulers below; make fix-copies is clean.

Scope: four changed, two deliberately not

Changed

Scheduler	Why
DDIMScheduler	Source of the bug in #12633 / #11347
DDIMParallelScheduler	Same formula in step, _get_variance, and batch path
CogVideoXDDIMScheduler	Same formula in step
CogVideoXDPMScheduler	Same formula in step

None of those step methods was a # Copied from of another; the formula was independently duplicated. The new helper is the shared copy point.

Not changed

Scheduler	Why list lookup is wrong there
PNDMScheduler	step_prk uses half-stride intermediates that are not entries of self.timesteps. step_plms rewrites prev_timestep and timestep when counter == 1. With skip_prk_steps, the list contains deliberate duplicates, so a first-match index would pick the wrong row.
RePaintScheduler	The schedule contains forward jumps. The next list entry is often a larger t, while the stride formula always means one reverse diffusion step. List lookup would break jump steps.

Off-schedule fallback

When step is called with a timestep that is not an entry of self.timesteps, previous_timestep falls back to the historical stride. That is not a defensive guess: replacing the fallback with a ValueError fails five existing tests that call step() outside the inference loop:

• test_eta
• test_inference_steps
• test_steps_offset
• test_time_indices
• test_timestep_spacing

Pipeline loops always pass schedule entries, so the linspace fix is not diluted by the fallback. Happy to convert the fallback to a raise and update those five tests if maintainers prefer a stricter API.

Issues

Closes #12633 and #11347. They are one root cause (stride vs linspace schedule), not two independent bugs.

Test plan

• test_prev_timestep_matches_schedule fails at pristine base 4b8e466bf and passes here
• Leading / trailing full-loop bit-identical to the old stride (max_abs_diff=0)
• Linspace prev sequence equals the schedule; trajectory changes
• tests/schedulers/test_scheduler_ddim.py and test_scheduler_ddim_parallel.py: 78 passed
• make style exit 0 (ruff + doc-builder + extra checks)
• make fix-copies exit 0

Platform: macOS arm64, CPU, CPython 3.12.

---

Self-review

Scope

Branch prev-timestep-schedule-mismatch against base 4b8e466bf.

Files:

• src/diffusers/schedulers/scheduling_ddim.py
• src/diffusers/schedulers/scheduling_ddim_parallel.py
• src/diffusers/schedulers/scheduling_ddim_cogvideox.py
• src/diffusers/schedulers/scheduling_dpm_cogvideox.py
• tests/schedulers/test_scheduler_ddim.py

Blocking issues

None.

Non-blocking issues

1. PNDM and RePaint still use the stride formula

   • Same line text appears, but list lookup is wrong for PLMS/PRK half-steps and RePaint jumps.
   • Leave for a follow-up if maintainers want a different design there.

2. Off-schedule step() fallback

   • Load-bearing for the five tests named above. Alternative is raise + update those tests; offered in the PR body.

Dead code

None introduced.

Copied-from

Block	Action
DDIMScheduler.previous_timestep	New source method
Parallel / CogVideoX DDIM / CogVideoX DPM previous_timestep	# Copied from DDIM; make fix-copies clean
step bodies	Not copies; each call site updated to use previous_timestep

Evidence

• test_prev_timestep_matches_schedule fails at base (AttributeError: previous_timestep) and passes here.
• Leading / trailing full-loop bit-identical (max_abs_diff=0); linspace changes (max_abs_diff≈2.58e2).
• DDIM default remains timestep_spacing="leading".
• make style and make fix-copies exit 0 on this branch (pre-existing tree-wide reformat noise from ruff format / doc-builder on unrelated files was not committed).

Verdict

READY.