MiniMax H3 gets two acceleration distillation LoRAs as PTD release sets a per-model record
linoy_tsaban · x · 2026-10-21
MiniMax H3's video model has now received two acceleration distillation LoRAs in quick succession: SMD (Simplified Matching Distillation) and a new PTD (Parametric Trajectory Distillation) release from taobao-content-ai, which the author notes may be a record for distillation finetunes per model.
PTD targets the Ref2VA (reference image → video + audio) task with a per-hop approach: each denoising step gets its own low-rank adapter (rank=128). At inference, each step selects its own adapter, so the 8-hop checkpoint runs in 8 NFEs and the 4-hop in 4. Weights are open on Hugging Face, with config embedded in safetensors metadata for automatic step derivation.
Related event: MiniMax H3 Gets Two Distillation LoRAs in One Week, Setting Record(2 posts)→
More from Multimodal
- Redditor remakes SD clown dance video at 25 steps with motion boost and fixed prompts — roychodraws · 2026-10-21
- Fibo-2, the #1 open image editing model, releases weights on Hugging Face — linoy_tsaban · 2026-10-21
- Ex-Brain Games host made a 25-min AI film in 5 days using 'vibe directing' in iMessage — lmoroney · 2026-10-21
- Motion Launches a General Video Agent That Edits Footage and Mimics Any Reference Style — mhdfaran · 2026-10-21
- MOSS-Image's 5B model stuns with legible-text movie posters that run locally — IcyPickle4540 · 2026-10-21
- What keeps AI videos watchable: hooks, consistency and a payoff — aftahi_ai · 2026-10-21