PDMD 4-NFE LoRA for MiniMax-H3 released on Hugging Face, distills video diffusion to 4 steps
linoy_tsaban · x · 2026-10-04
The PDMD team released a 4-step (4-NFE) distilled LoRA for MiniMax-H3 on Hugging Face, from the paper "PDMD: Projected Distribution Matching Distillation for Video Diffusion Models" (arXiv:2609.35768).
- LoRA targets the transformer (MiniMaxH3Transformer3DModel); VAE, audio VAE, schedulers and text encoder unchanged from the base model
- Specs: rank 128 / alpha 128 / bf16, 312 LoRA pairs (1.38GB); fusion via Wbase += lorascale (loraB @ loraA) with lorascale=1.0
- Sample with 4 denoising steps using the base scheduler config (shift 12/3)
- Apache-2.0; a full transformer-weights version (pdmd4NFEfull) is also available
Community testing: impressive quality for 2-4 steps, though close-ups and speech still trail H3 Turbo.
Related event: Two-Step PDMD LoRA for MiniMax-H3 Released Open-Source(2 posts)→
More from Multimodal
- Cloudflare's new vision model clef appears on Hugging Face as devs urge Perplexity to add it — ostrisai · 2026-10-04
- Sopro V2 Turbo 2610: cleaner voice cloning, 120M model, ~300ms first audio on CPU — SammyDaBeast · 2026-10-04
- 2-step PDMD H3 LoRA tested: impressive for 2 steps, but H3 Turbo wins close-ups and speech — linoy_tsaban · 2026-10-04
- AI-generated Japanese town spawns unprompted interiors, including an art studio no one asked for — Merzmensch · 2026-10-04
- Handing off AI-generated images as decomposed RGBA layers for designer edits — tinkerbellyie · 2026-10-04
- Fan uses AI to put Asmongold into World of Warcraft and plays him — djcows · 2026-10-04