MiniMax-H3 PDMD 2-step LoRA in ComfyUI: bf16 file, strength 1.0, 2 steps, no CFG

This repository converts the PDMD 2-NFE LoRA, the 2-step student distilled from MiniMax-H3 33B with Projected Distribution Matching Distillation (rank 128 on the attention projections and both feed-forward layers of all 50 transformer blocks plus 2 token-refiner blocks), from diffusers PEFT keys to the ComfyUI H3 layout. It is the companion of the 4-NFE conversion Iwannapose/minimax_h3_pdmd_4nfe_comfyui. Files: minimax_h3_pdmd_2nfe_comfyui.safetensors (bf16, recommended, loads in the stock loader) and a numerically identical fp32 copy with no benefit on GPU. Both are larger than the source because the fused qkv_proj stores its block-diagonal zero padding. Conversion: q/k/v are fused into qkv_proj (A stacked by rows, B block-diagonal), the two 14336-row halves of each mlp.fc1 lora_B are swapped because diffusers SwiGLU outputs [value; gate] while ComfyUI H3 expects [gate; up], alphas are set so alpha/rank is 1.0 (the PDMD fusion scale), and modules are renamed to ComfyUI names. The author reports 468 numeric checks passing against the originals (exact in fp32), a key set identical to a known-good turbo LoRA, all 208 target keys present in the H3 ref2va int8 model, and a tensor-for-tensor match with an independent converter. Use: LoraLoaderModelOnly (the H3 text encoder loads separately), strength exactly 1.0, since other values rescale the distillation delta; sample 2 denoising steps with the H3 scheduler (shift 12 video, 3 audio) and no CFG, as H3 is guidance-distilled. Licence on the card: Apache-2.0; MiniMax-H3 keeps its own terms. Based on the Iwannapose/minimax_h3_pdmd_2nfe_comfyui card pinned to revision 3acb11d6f9bc65e68bb907a5c6b596cd719c4360 and observed on 2026-10-27. Verification claims are the converter's; quality of 2-step output was not checked. Nothing was run. 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/Iwannapose/minimax_h3_pdmd_2nfe_comfyui/raw/3acb11d6f9bc65e68bb907a5c6b596cd719c4360/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 241:659, 704:956, 2553:2946, 2649:2946, 2794:2946, 2883:2946, 2183:2531; exact excerpt SHA256 502411d4dd318a2469267edce6825145e224afdd2177e96b5e5c56303d0cc1bc,76d4bda3ce7f4cb3094e7636ccbd698fe860fd896f3881e4da6d24a549883034,423813e89d14bfc76be3441a4917d575967c9e0494397ab32ea000"
    }
  ]
}