MiniMax-H3 PDD baked checkpoints for ComfyUI: with or without heads, and the 8-step recipe

aptech0081 converts alibaba-pai's MiniMax-H3 Parallel Decoding Distillation (PDD) weights into ComfyUI files; these are format conversions, not new training. The Ref2VA checkpoints generate video and audio from reference conditioning. Which file: minimax_h3_ref2va_pruned_pdd_acc_8step_baked_with_heads_int8_convrot (21.06 GB, about 38 percent smaller, built on the official Comfy-Org pruned Ref2VA INT8 base) is the recommended native choice; minimax_h3_ref2va_pdd_acc_8step_baked_with_heads_int8_convrot (34.12 GB) is the full version. Both carry the merged PDD trunk plus 32 video and 32 audio output heads and load in native ComfyUI after PR #17157 (merged 18 October 2026). The without-heads checkpoint (formerly ..._baked_int8_convrot, same bytes) keeps the original single output heads and still needs the PDD head bank and Apply node. Two ..._comfyui.safetensors files (FL2VA and Ref2VA) hold a rank-64 trunk LoRA and head bank for the companion node pack. Baked checkpoints go in ComfyUI/models/diffusion_models/; the LoRA files in ComfyUI/models/pdd_acc/. Native recipe (with heads): update ComfyUI to include PR #17157, load the checkpoint with Load Diffusion Model (UNETLoader), use the usual MiniMax H3 Ref2VA conditioning, text encoder, video VAE and audio VAE, then Euler, simple scheduler, 8 steps, CFG 1.0, denoise 1.0, with the default video/audio sigma shifts 12.0/3.0. Four steps also work; start with eight. Do not apply the PDD LoRA or Apply node again and keep other acceleration LoRAs off. config.transformer.pdd_steps = 32 is the head-bank size, not the sampler step count. Without heads (ComfyUI-MiniMax-H3-PDD-Acc nodes): lora_strength 0.0 (trunk already baked), head_strength 1.0, nfe 8 (or 4), Euler, CFG 1.0, shifts 12/3, and connect the Apply node's sigma output to SamplerCustomAdvanced. Standalone LoRAs: trunk and head strengths 1.0 on the stock FL2VA or Ref2VA checkpoint. Caveats: the pruned file was dequantized, merged and requantized and uses the pruned timing curve, so results can differ from the full file (sampled requantization error up to about 0.74 percent; timing curve error about 1.4e-5); the full with-heads file keeps the trunk bytes and adds float32 heads. File size is not total VRAM, and no minimum VRAM is claimed. Licence: MiniMax H3 Community License Agreement. Based on the aptech0081/MiniMax-H3-PDD-ComfyUI card pinned to revision c169e9cd1dab4ac56b413b1343218ab4a3af2f19 and observed on 2026-10-27. Nothing was run here; the validation reports, conversion script and ComfyUI PR were not inspected, and no rendered-quality comparison exists in the card. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/aptech0081/MiniMax-H3-PDD-ComfyUI/raw/c169e9cd1dab4ac56b413b1343218ab4a3af2f19/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 604:1042, 1882:2287, 3019:3356, 3189:3356, 3594:3685, 5371:5527, 6939:7169; exact excerpt SHA256 9bb132a8f66524bec5bc0e3157003d78e468cd2ef70cdda7547a1a8efd40f9fc,3ad233c7cda1bc45c245ad530d6ede35a6dde4202a82fd28e32ca10eaa638b8d,685ff8d6161513fe3d9d1999221be10eff0fe9217c11d5486a7"
    }
  ]
}