Fizgig 9.8.3 blends MiniMax H3 Image and video block-by-block to fix frozen fast action
shootthesound · reddit · 2026-11-22
Fizgig 9.8.3 adds an experimental Frankenstein mode: with MiniMax H3 Image loaded in Studio, per-block sliders set how much of each block comes from H3 video, rendered side-by-side with plain H3 Image for comparison.
- Motivation: H3 video handles motion better; H3 Image tends to render fast action frozen and sharp.
- After testing every block across action scenes, motion sits roughly in blocks 26-36; the Motion preset (blocks 27-28 and 32-34 at 0.65) carries motion blur, spray and smoke while keeping H3 Image's look. Blending too many blocks or too strongly breaks the image.
- The demo pair shows hand position, tucked leg and floor clearance represented better in the blend in over 90% of tests (20 total).
- Blends can be saved as a new H3 Image base model that loads in ComfyUI like the original. Requires 16GB+ VRAM and 48GB RAM (64GB with a 16GB card).
More from Multimodal
- ViewNeXt world model flies over any scene from one image, tops LPIPS vs 15 methods — janusch_patas · 2026-11-22
- Creature shaders: materials self-sort and grow into organs to spawn new species — round · 2026-11-22
- fofrAI breaks down the workflow behind his vibe-driven AI images, with copyable agent instructions — fofrAI · 2026-11-22
- Lakonik's pi-H3 on Hugging Face Delivers Ultra-Fast Generation in Seconds — Friendly-Fig-6015 · 2026-11-22
- Breaking the world model amnesia window: GAE with 81-frame memory revisits scenes at high PSNR — jonstephens85 · 2026-11-22
- Gemini Neon generates a stunning interactive guide to the Riemann Hypothesis — intellectronica · 2026-11-22