[MiniMax-H3] Load the pruned ComfyUI checkpoints and add single-file support for MiniMax-H3-Image

Ouverte
#15,302 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

@guoriyue y travaille déjà.

Depuis le 15/11/2026.

  • #15299 par @guoriyue — ouverte

Évaluation

Difficulté
4/5
Temps estimé
3-5 jours
Accessibilité débutants
28/100
Type d'issue
Fonctionnalité
Clarté
Plutôt claire
Activité
Active
Stack technique
python, pytorch

Piste de recherche

The referenced implementation is PR #15299, already open, so start by reading that diff rather than the issue text alone. The transformers live under src/diffusers/models/transformers/ (MiniMaxH3Transformer3DModel, MiniMaxH3ImageTransformer2DModel), and the single-file mapping is in src/diffusers/loaders/single_file_model.py plus the MiniMax-H3 original-config file in single_file_utils.py. The open design question — whether a new config_from_checkpoint_fn key in the single-file registry is acceptable — needs a maintainer answer before any patch, and done means a pruned ComfyUI checkpoint loads through from_single_file with tests covering the unstripped config path.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

Is your feature request related to a problem? Please describe.

ComfyUI distributes pruned MiniMax-H3 checkpoints (Comfy-Org/MiniMax-H3, Comfy-Org/MiniMax-H3-Image). They cannot be loaded in diffusers today:

  • The pruned files replace the timestep MLP and the wide AdaLN projections (about 40% of the parameters) with the activated timestep curve sampled on a 1025-point grid (adaln_t_table) and 8-input AdaLN projections. MiniMaxH3Transformer3DModel has no module that matches these weights, and from_single_file refuses adaln_t_table.
  • MiniMaxH3ImageTransformer2DModel has no single-file support at all (no FromOriginalModelMixin), and the ComfyUI file stores the PDD heads as a stacked bank (head 0 plus per-step offsets) rather than as pdd_proj_out.{i}.

Describe the solution you'd like.

  1. An opt-in config field (adaln_curve_grid) on both transformers. When set, time_embedder is a lookup of the sampled curve, linearly interpolated as ComfyUI does, and the AdaLN projections skip their SiLU. Stock configs keep their current behaviour.
  2. from_single_file for both transformers: map adaln_t_table to the table, unstack the PDD head bank for the image model, and refuse quantized ComfyUI files with a message naming the bf16 file.
transformer = MiniMaxH3ImageTransformer2DModel.from_single_file(
    "https://huggingface.co/Comfy-Org/MiniMax-H3-Image/blob/main/diffusion_models/minimax_h3_image_pruned_bf16.safetensors",
    dtype=torch.bfloat16,
)

Design question for maintainers

The repository config on the Hub describes the unpruned model, and no Hub repo carries a pruned diffusers config that a model type could point at. So the pruned shape (adaln_curve_grid, time_embed_dim) has to be read off the checkpoint itself. #15299 does this with a new optional key in the single-file registry, config_from_checkpoint_fn. Is a new registry key acceptable here, or would you prefer another route (for example a separate model type with its own config)?

Additional context.

Implementation proposed in #15299.

Langage dominant
Python
Étoiles
34.9k
Forks
7.5k
Merge moyen
3 j 16 h
PR mergées (30 j)
65

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de huggingface/diffusers

Toutes les issues de huggingface/diffusers

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.