Skip to content

feat(mm): recognize MiniMax H3 Ref2VA models - #161

Open
lstein wants to merge 3 commits into
mainfrom
feat/minimax-h3-ref2va-recognition
Open

feat(mm): recognize MiniMax H3 Ref2VA models#161
lstein wants to merge 3 commits into
mainfrom
feat/minimax-h3-ref2va-recognition

Conversation

@lstein

@lstein lstein commented Aug 27, 2026

Copy link
Copy Markdown
Collaborator

Summary

First of three stacked PRs adding MiniMax H3 Ref2VA (reference-conditioned video+audio) support. This one makes Ref2VA models recognized, installable, and correctly labeled — generation support follows in the next PR.

  • Taxonomy: MiniMaxH3VariantType.REF2VA (HF subfolder transformer_ref/; same architecture as FL2VA, different weights, reference task only).
  • Identification: the hard rejection of ref2va filenames becomes filename-based variant detection (the two task transformers are key-for-key indistinguishable — filename is the classifier, the variant override corrects renamed files). A transformer_ref/-only diffusers folder identifies as ref2va and is always marked components_only (its folder weights are not loadable in v1 — the generation path is a components install plus a single-file transformer override); a dual-transformer folder stays fl2va.
  • Starter models: MiniMax H3 Ref2VA Transformer (int8, pruned) (~21 GB) and the MiniMax H3 Ref2V Turbo LoRA (4-step distillation) join the bundle.
  • Fixtures: three new identification fixtures (ref2va single file, transformer_ref/-only slim folder, variant-override case).
  • Legacy frontend: zod enum + label; openapi.json/schema.ts regenerated.
  • webv2: variant label/options; a temporary transformer-picker gate (removed in the webv2 PR) so an installed Ref2VA checkpoint cannot silently generate under FL2VA's config; the Turbo accelerator auto-pick excludes ref2v-token LoRAs so the Ref2VA distillation never auto-applies to FL2VA runs.
  • Model loader invocation: fails fast on a Ref2VA transformer (replaced with variant stamping in the next PR).

Testing

  • 85 identification tests (3 new fixtures), starter-model tests, loader-node tests, 1228 model-manager tests — all green.
  • webv2 taxonomy/videoPolicies suites green; both frontends typecheck.
  • Adversarial fresh-context review completed; all confirmed findings fixed in this PR.

Stack

  1. → this PR — recognition
  2. feat(backend): MiniMax H3 Ref2VA generation — packing, conditioning, invocations #162 — backend generation (numerics port + nodes)
  3. feat(webv2): MiniMax H3 Ref2VA reference mode #163 — webv2 reference mode

🤖 Generated with Claude Code

https://claude.ai/code/session_01S4B5exWsbC2z2Uu2tA167A

Adds the ref2va variant to MiniMaxH3VariantType and teaches identification
to classify it:

- Single-file checkpoints: the former hard rejection of ref2va filenames
  becomes filename-based variant detection (the two task transformers are
  key-for-key indistinguishable, so the filename is the classifier and the
  variant override field corrects renamed files).
- Diffusers folders: a transformer_ref/-only folder now identifies as
  ref2va; a dual-transformer folder stays fl2va. Ref2VA folder weights are
  not folder-loadable in v1, so a ref2va folder is always marked
  components_only - the generation path is a components install plus a
  single-file transformer override.
- Starter models: Ref2VA pruned int8 transformer (~21 GB) and the Ref2V
  4-step Turbo LoRA join the MiniMax H3 bundle.
- Identification fixtures: ref2va single file, transformer_ref-only slim
  folder, and a variant-override case.
- Legacy frontend: ref2va in the zod variant enum + label map; openapi.json
  and schema.ts regenerated.
- webv2: variant label/options, a temporary transformer-picker gate, and a
  ref2v-token exclusion in the Turbo accelerator auto-pick so the Ref2VA
  LoRA never auto-applies to FL2VA runs.
- Model loader invocation: fails fast on a Ref2VA transformer (override or
  main) so hand-authored workflows cannot silently generate degraded
  output; TODO(ref2va) markers note what the follow-up removes.

Generation support (packing, reference conditioning, UI mode) lands in
follow-up PRs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S4B5exWsbC2z2Uu2tA167A
lstein added 2 commits August 29, 2026 14:38
…a-recognition

# Conflicts:
#	invokeai/frontend/webv2/src/features/video/core/videoPolicies.ts
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant