Skip to content

Models: MiniMax-H3-Image - #72

Merged
Nynxz merged 5 commits into
mainfrom
h3-image
Nov 12, 2026
Merged

Nynxz merged 5 commits into
mainfrom
h3-image

Conversation

@Nynxz

@Nynxz Nynxz commented Nov 12, 2026 •

Copy link
Copy Markdown
Owner

MiniMax-H3-Image (MiniMax's 6-step, CFG-free image build of H3) in the model list, with a recipe.

What's in it

  • ComfyUI: pin v0.45.0. The model code and template first ship there.
  • Models: read ComfyUI v0.45's templates right.
    • A text encoder that only feeds TextGenerate* is a prompt rewriter, told apart by its wiring rather than its file name. MiniMax-H3-Image's rewriter is the same file as MOSS Image's text encoder, so the name can't tell them apart. ERNIE-Image and HiDream O1 templates benefit too.
    • A new template_skip list in the catalog leaves out templates that load another family's diffusion model with their own encoder and VAE. One v0.45 template loads H3's diffusion model this way, which put its encoder and VAE in H3's lists.
  • Web: offer to download a recipe's training adapter. When a recipe names an adapter that isn't downloaded, the picker says "Not downloaded yet · Download 155 MB" instead of showing a bare file name.
  • Models: MiniMax-H3-Image.
    • The model has one output head per step of its own 6-level noise grid. Training draws only those levels, each through its own head; the patch in compat/minimax_h3.py gives the model its grid while training, when there's no sampler schedule.
    • Training runs over Ostris's de-distillation adapter, which the recipe picks and which is now in the adapter list.
    • Samples are 2048², 6 steps, CFG 1, simple scheduler, as ComfyUI's template uses.
    • Added to the catalog's tested list.
  • Web: the new-run form fits a phone with a recipe picked. This bug predates this branch. Below 1000px the layout's single column was a plain 1fr, so a recipe's one-line summary pushed the form to about 900px on a 390px screen. It's now minmax(0, 1fr), as the two-column layout already uses.

Training results

The profile's defaults leave the learning rate at our default 1e-4, which learns weakly on this model. A note in the profile gives what we measured:

  • Face set, 37 pictures: at 1e-4 it barely learned. At 1e-3 with gradient accumulation 4, spread noise levels and a low-noise lean of 1.5, likeness reached 0.66 mean (buffalo_l) at about 2000 steps.
  • Style set, 8 pictures:
    • at 1e-3 and 5e-4 it overcooked within a few hundred steps: blotchy texture, and the dataset's characters leaking into unrelated prompts;
    • at 2e-4 it stayed clean for 750 to 1000 steps.

Tests

  • Engine tests: pass, apart from test_sharpest_frame_picks_the_detailed_one (already failing on main). Two harness tests also fail locally because the machine had no ComfyUI folder_paths module; they fail the same way without these changes.
  • tests/comfy on ComfyUI v0.45.0: 8/8 pass, including the new h3image family, whose check confirms the patched model computes what ComfyUI's code does.
  • Web: vue-tsc --noEmit and vite build pass. Checked the New Run page at 390px.

🤖 Generated with Claude Code

Summary by CodeRabbit

  • New Features
    • Added a MiniMax-H3-Image training recipe and model listing, with setup guidance and a compatible adapter option.
    • The adapter picker now identifies recommended adapters that aren’t installed and offers a download action.
    • Improved template selection to distinguish prompt-rewriting text encoders from image-generation encoders.
  • Bug Fixes
    • Improved the New Run page layout on narrower screens.

Nynxz and others added 5 commits November 12, 2026 14:18
MiniMax-H3-Image's model code and template arrive in v0.45.0.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A text encoder that only feeds TextGenerate is a prompt rewriter, told apart by its wiring rather than its file name (MiniMax-H3-Image's rewriter is MOSS Image's text encoder). Templates that load another family's diffusion model with their own encoder and VAE can be left out (template_skip), so they don't add to that family's lists.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A recipe can name an adapter that isn't downloaded yet; the picker now says so and offers the download instead of showing a bare file name.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Training draws the model's six sampling noise levels, each through its own output head, over Ostris's de-distillation adapter (now in the adapter list). Samples at 2048², 6 steps, CFG 1. The profile notes the learning rates measured on a face set and a style set. Smoke-tested in tests/comfy.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On narrow screens the layout's single column was a plain 1fr, which grows to its content's min-content width: a recipe's one-line summary pushed the form to about 900px. minmax(0, 1fr), as the two-column layout already uses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@coderabbitai

coderabbitai Bot commented Nov 12, 2026 •

Copy link
Copy Markdown

Explain Change →

No actionable comments were generated in the recent review.

Walkthrough

The pull request adds MiniMax-H3-Image recipe and compatibility support, changes template and adapter discovery, and adjusts the New run page’s narrow-screen layout.

Changes: MiniMax H3 Image support

Layer / File(s) Summary
Recipe and adapter setup
engine/profiles/minimax-h3-image.json, engine/catalog/families.json, engine/zentrainer/models/adapters.py, web/src/components/AdapterPicker.vue, CHANGELOG.md
Adds the MiniMax-H3-Image profile and adapter catalog entry. The adapter picker can show and offer a matching adapter that is not installed. The catalog marks the family as tested.
H3 training compatibility and smoke tests
engine/zentrainer/harness/compat/minimax_h3.py, tests/comfy/smoke.py, tests/comfy/test_training.py
The compatibility wrapper supplies the H3 Image sigma grid as sample_sigmas during training when that option is absent. The smoke tests add the H3 Image family, pass model metadata through loading, and include the family in training tests.
Template and encoder discovery
engine/catalog/families.json, engine/zentrainer/models/families.py, tests/engine/test_families.py, CHANGELOG.md
Family discovery applies configured template-name exclusions. Template parsing excludes loaders used only by text-generation nodes from text-encoder selection. Tests check template inclusion and H3 Image text-encoder and VAE assignments.

Narrow-screen layout

Layer / File(s) Summary
New run narrow-screen layout
web/src/views/NewRunView.vue, CHANGELOG.md
At viewport widths up to 1000px, the single grid column uses minmax(0, 1fr).

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Feature

Possibly related PRs

  • Nynxz/ZenTrainer#40: Updates the same Unreleased changelog section; merge order affects where these changelog entries appear.
Merge Risk: Minimal · up to 6c18c

This change adds the MiniMax-H3-Image recipe, adapter download support, and a narrow-screen layout fix. No concrete merge-blocking issue was found.

Pre-merge checks: Passed 4

Check name Status Explanation
Description Check Passed Check skipped - CodeRabbit’s high-level summary is enabled.
Title check Passed The title "Models: MiniMax-H3-Image" names the main change. The PR adds the MiniMax-H3-Image model to the model list with a training recipe, profile, and compatibility patches. The title is short and …
Linked Issues check Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check Passed Check skipped because no linked issues were found for this pull request.
Full details: Title check

Explanation

The title "Models: MiniMax-H3-Image" names the main change. The PR adds the MiniMax-H3-Image model to the model list with a training recipe, profile, and compatibility patches. The title is short and specific. It omits the supporting work (template handling, adapter download, layout fix), which the title does not need to cover.

Finishing Touches

Generate unit tests (beta)

  • Commit to this branch
  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

Recent review info

Run configuration

Commits

Reviewing files that changed from the base of the PR and between 99ff528 and 6c18c65.

Files ignored due to path filters: 1

  • comfy.lock is excluded by !**/*.lock, !comfy.lock
Files selected for processing: 11

  • CHANGELOG.md
  • engine/catalog/families.json
  • engine/profiles/minimax-h3-image.json
  • engine/zentrainer/harness/compat/minimax_h3.py
  • engine/zentrainer/models/adapters.py
  • engine/zentrainer/models/families.py
  • tests/comfy/smoke.py
  • tests/comfy/test_training.py
  • tests/engine/test_families.py
  • web/src/components/AdapterPicker.vue
  • web/src/views/NewRunView.vue

Included review availability: This review used your included allowance. Your plan provides up to 1 included review per hour; 0 remain after this review.

@Nynxz
Nynxz merged commit e9d2ac0 into main Nov 12, 2026
4 checks passed
@Nynxz
Nynxz deleted the h3-image branch November 12, 2026 04:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant