ProMoE — Hub custom pipeline

Load checkpoints with native Hugging Face diffusers and this folder on the Hub (or via custom_pipeline):

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/ProMoE-diffusers",
    trust_remote_code=True,
    torch_dtype=torch.float16,
)
pipe.to("cuda")

Hub layout

ImageNet class labels

Each variant keeps an English id2label map in model_index.json (DiT-style).

• pipe.id2label — id → English label (comma-separated synonyms)
• pipe(class_labels=207, ...) — class-conditional sampling with integer ids

Copy the full 1000-class id2label block from BiliSakura/DiT-diffusers when publishing a model repo.

model_index.json

Copy entries from model_index.json.example into your model repo after save_pretrained. Use ["_class_name"] = ["pipeline", "ProMoEPipeline"] and custom module stems for each component.

• FlowMatch scheduler: "scheduler": ["scheduling_flow_match_promoe", "ProMoEFlowMatchScheduler"]
• VAE: "vae": ["diffusers", "AutoencoderKL"] with stabilityai/sd-vae-ft-mse weights or bundled safetensors
• ProMoE-TC presets: ProMoE_TC_S, ProMoE_TC_B, ProMoE_TC_L, ProMoE_TC_XL (see convert script)

Regenerate: python scripts/build_community_pipelines.py