pi-H3 distillation test: 4 steps passable, 2 steps break motion and audio
BittiAI · reddit · 2026-11-22
A Reddit user tests pi-H3 video generation with 4-step and 2-step inference (no VAE decoder). Four steps: quality middling but some decent shots, prompt following and motion largely intact, audio bad as expected, weak reference following, pixelated transitions from a code issue. Two steps: essentially unusable — garbage audio, incorrect motion, marginal prompt following. The attached frame took 3m20s to generate at 1344×768 on an RTX 5090.
More from Multimodal
- Open-source Short-Drama Skill helped a creator ship 20+ AI dramas with 50M+ views — vista8 · 2026-11-22
- A burger menu designed as an arcade game screen, prices denominated in HP — gizakdag · 2026-11-22
- AI-generated short film: an AI actress starts questioning her world — THEBLUEHORIZON_AI · 2026-11-22
- Kling Image 4.0 vs ChatGPT Images 2.6: group portrait test gives Kling a slight edge — iamfakhrealam · 2026-11-22
- Kling Image 4.0 vs GPT Image 2.6: same prompt, too close to call — JaynitMakwana · 2026-11-22
- LATO.2 open-sourced: image-to-low-poly 3D mesh with controllable vertex counts, runs on 8GB VRAM — Promptmethus · 2026-11-22