Skip to content

Fix HiDiffusion ratio overrides and SDXL T2 behavior - #9511

Open
DustyShoe wants to merge 10 commits into
invoke-ai:mainfrom
DustyShoe:fix/hidiffusion-ratio-behavior
Open

Fix HiDiffusion ratio overrides and SDXL T2 behavior#9511
DustyShoe wants to merge 10 commits into
invoke-ai:mainfrom
DustyShoe:fix/hidiffusion-ratio-behavior

Conversation

@DustyShoe

@DustyShoe DustyShoe commented Aug 16, 2026

Copy link
Copy Markdown
Collaborator

Summary

This fix makes HiDiffusion T1 and T2 ratio overrides behave as configured.

For standard high-resolution SDXL generation, the aggressive RAU-Net schedule replaced the T2 threshold with a fixed step, so changing T2 did not affect the result. The aggressive schedule is now limited to T1 blocks, while T2 blocks use their configured threshold.

Ratio overrides are stored per patched model instead of mutating process-global preset dictionaries. This prevents overlapping generations from leaking settings into each other. The UI now supports automatic ratios, which preserve HiDiffusion's resolution-aware defaults, while explicit values remain available as manual overrides.

Regional prompt and regional IP-Adapter masks now include the additional 16x downscale level used by HiDiffusion's RAU-Net. This prevents missing-mask KeyError failures when Regional Guidance is combined with HiDiffusion.

The Redux migration converts the previous numeric defaults to automatic ratios, metadata recall preserves automatic values, and the OpenAPI schema, generated frontend types, documentation, and localized UI copy are updated.

Related Issues / Discussions

None.

QA Instructions

  1. Generate the same SDXL image twice with HiDiffusion and RAU-Net enabled, using T2 values of 0 and 1. Confirm that the outputs differ.
  2. Enable Automatic Ratios and generate at both standard high resolution and extreme resolution. Confirm that the resolution-aware presets are used.
  3. Disable Automatic Ratios and confirm that manual T1 and T2 values are included in the denoise node.
  4. Recall metadata from an image generated with Automatic Ratios and confirm that automatic mode is restored.
  5. Generate with HiDiffusion and Regional Guidance enabled. Confirm that denoising completes without a missing regional mask error.

Automated validation:

  • Backend stable diffusion tests: 28 passed.
  • Frontend migration, metadata, and graph tests: 164 passed.
  • TypeScript typecheck passed.
  • Ruff check passed.
  • Prettier check passed.
  • git diff --check passed.

Merge Plan

No special merge handling is required.

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)

@github-actions github-actions Bot added python PRs that change python files invocations PRs that change invocations backend PRs that change backend files frontend PRs that change frontend files python-tests PRs that change python tests docs PRs that change docs labels Aug 16, 2026
@lstein lstein added the 6.14.1 label Aug 17, 2026
@lstein lstein moved this to 6.14.x Theme: USER EXPERIENCE in Invoke - Community Roadmap Aug 17, 2026
@lstein lstein moved this from 6.14.x Theme: USER EXPERIENCE to 6.14.1: Bug fixes to 6.14.0 in Invoke - Community Roadmap Aug 17, 2026
@DustyShoe
DustyShoe marked this pull request as ready for review August 20, 2026 14:09
@DustyShoe
DustyShoe force-pushed the fix/hidiffusion-ratio-behavior branch from cc121ba to c9eb228 Compare August 24, 2026 08:32
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

6.14.1 backend PRs that change backend files docs PRs that change docs frontend PRs that change frontend files invocations PRs that change invocations python PRs that change python files python-tests PRs that change python tests

Projects

Status: 6.14.1: Bug fixes to 6.14.0

Development

Successfully merging this pull request may close these issues.

3 participants