feat(model): Add ER SDE / DPM++ 2M Scheduler Support For Anima

[kappacommit]

Summary

ER SDE is the primary recommended scheduler for Anima. After testing, it produces the best and most coherent outputs I've seen for the model.

Structurally, the easiest way to add support for ER SDE for Anima, was to also add support for SD (1.5, 2, SDXL). So this PR adds support for Anima and SDXL.

ER SDE is not available in diffusers, so a custom implementation was required.

Reference implementation was here: https://github.com/QinpengCui/ER-SDE-Solver this is where all of the fancy math and logic comes from (MIT licensed, free to use)

I also add support for DPM++ 2M Scheduler

Related Issues / Discussions

Requested in discord https://discord.com/channels/1020123559063990373/1149506274971631688/1501016017009246379

QA Instructions

• Use any SDXL model
• Select the ER SDE scheduler
• Generate an image successfully

--

• Use any Anima model
• Select the ER SDE scheduler
• Generate an image successfully

--

• Use any Anima model
• Select the DPM++ 2M SDE scheduler
• Generate an image successfully

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[lstein]

@JPPhoto I've assigned this to you and labeled it for 6.13.0 because it does fit the theme of model support enhancements. But let me know if you think it is too late to get it in, and we'll go for 6.14.

[JPPhoto]

@JPPhoto I've assigned this to you and labeled it for 6.13.0 because it does fit the theme of model support enhancements. But let me know if you think it is too late to get it in, and we'll go for 6.14.

Let's see if we can get it in, but if timing is too tight then this shouldn't block the release.

[JPPhoto]

• invokeai/app/invocations/anima_denoise.py:506 - Anima dpmpp_2m and dpmpp_2m_sde ignore clipped img2img/inpaint sigma schedules. denoising_start / denoising_end slice sigmas, and initial latents are noised with that clipped sigmas[0], but these DPM++ schedulers fall back to set_timesteps(num_inference_steps=total_steps) because Diffusers does not accept sigmas= there. That regenerates a full schedule starting near sigma 1.0. Example: steps=30, denoising_start=0.5 should start at sigma 0.75, but the scheduler starts around 0.9997. To expose this issue, add a test that sets up Anima DPM++ with denoising_start=0.5 and asserts the scheduler's first sigma/timestep matches the clipped schedule.
• Test gap: invokeai/tests/backend/flux/test_anima_schedulers.py:80 - The DPM++ schedule equivalence test only covers a full schedule. It misses denoising_start and denoising_end, which is where the runtime path diverges.
• Test gap: invokeai/tests/app/invocations/test_anima_denoise_er_sde_dispatch.py:8 - The test comment says numerical correctness is gated by a parity test against er_sde_rf_step, but that parity test is not present in this PR. The included ER-SDE tests are mostly construction, finite-output, and structural tests, so they do not prove parity with the previous Anima ER-SDE implementation.
• Test gap: invokeai/backend/stable_diffusion/schedulers/schedulers.py:95 - er_sde is added to the general SD/SDXL scheduler map, but coverage is only synthetic VP smoke testing. There is no integration test through get_scheduler() or a real SD/SDXL denoise path proving config loading, prediction type override, and scheduler stepping work together.

[kappacommit]

• invokeai/app/invocations/anima_denoise.py:506 - Anima dpmpp_2m and dpmpp_2m_sde ignore clipped img2img/inpaint sigma schedules. denoising_start / denoising_end slice sigmas, and initial latents are noised with that clipped sigmas[0], but these DPM++ schedulers fall back to set_timesteps(num_inference_steps=total_steps) because Diffusers does not accept sigmas= there. That regenerates a full schedule starting near sigma 1.0. Example: steps=30, denoising_start=0.5 should start at sigma 0.75, but the scheduler starts around 0.9997. To expose this issue, add a test that sets up Anima DPM++ with denoising_start=0.5 and asserts the scheduler's first sigma/timestep matches the clipped schedule.

  • Test gap: invokeai/tests/backend/flux/test_anima_schedulers.py:80 - The DPM++ schedule equivalence test only covers a full schedule. It misses denoising_start and denoising_end, which is where the runtime path diverges.

  • Test gap: invokeai/tests/app/invocations/test_anima_denoise_er_sde_dispatch.py:8 - The test comment says numerical correctness is gated by a parity test against er_sde_rf_step, but that parity test is not present in this PR. The included ER-SDE tests are mostly construction, finite-output, and structural tests, so they do not prove parity with the previous Anima ER-SDE implementation.

  • Test gap: invokeai/backend/stable_diffusion/schedulers/schedulers.py:95 - er_sde is added to the general SD/SDXL scheduler map, but coverage is only synthetic VP smoke testing. There is no integration test through get_scheduler() or a real SD/SDXL denoise path proving config loading, prediction type override, and scheduler stepping work together.

• Feedback 1+2 (DPM++ img2img bug): Fixed and added a passing test covering the clipped-schedule case.

• Feedback 3 (false parity claim): This was just a comment issue. It reference a previous implementation of er_sde, which never existed. Updated the comment, but kept the test because it still has value.

• Feedback 4 (SD/SDXL integration test): Not addressed — inconsistent with the rest of the codebase where no other
  scheduler has a stepping integration test.