8-step and 4-step PTD acceleration LoRAs released for MiniMax-H3 Ref2VA
Apprehensive_Sky892 · reddit · 2026-10-22
A community member (unaffiliated with the project) released 8-step and 4-step acceleration LoRAs for MiniMax-H3's Ref2VA (reference-image → video + audio) task, distilled via Parametric Trajectory Distillation and hosted on Hugging Face (taobao-content-ai/MiniMax-H3-PTD-LoRAs).
Unlike a single shared LoRA, PTD trains one independent per-hop LoRA per inference step on top of PDD's per-interval output heads; at inference each denoising step selects its own adapter, so an 8-hop checkpoint runs in 8 NFE and 4-hop in 4. ComfyUI compatibility with the standard LoRA loader is unconfirmed.
Related event: MiniMax H3 Gets Two Acceleration LoRAs in One Week, a Single-Model Record(3 posts)→
More from Multimodal
- Non-coder builds AI filmmaking studio with Suno, Midjourney, Kling and Claude to enter $50k contest — PatriciaCoelhoV · 2026-10-22
- fal employee shows Claude + fal generating a game in 60 minutes — OdinLovis · 2026-10-22
- Indie dev uses AI to 1:1 recreate his own RV for his fourth game — ezshine · 2026-10-22
- User Shows Off Cinematic AI Video Made with Seedance 2.5 — SimplyAnnisa · 2026-10-22
- Fudan's TextRIRO enables scene text editing with self-paired real data only — Fudan-University · 2026-10-22
- StreamEval benchmarks streaming video generation: best model scores just 36.90 on task execution — StreamEval Team · 2026-10-22