MiniMax H3 image-trained LoRA tested for video generation
bdsqlsz · x · 2026-08-16
Kohyass shared a technical practice demonstrating that a LoRA trained on images generated by MiniMax H3 can also be applied to video generation.
Experiment Setup:
- Data: 20 images of an original character
- Training steps: 300 steps
- Key parameters: --h3guidancelossscale 4, --h3guidancelosssigmamin 0.15
A demo video (with audio) was attached to show the results.
More from Multimodal
- A Midjourney style reference worth trying: --sref 3623850304 — umesh_ai · 2026-08-16
- MiniMax H3 achieves breakthrough character consistency in video generation — Aiden_Tech_Ai · 2026-08-16
- Video 'Bang Bang' Created with Grok Imagine and Topaz — creatoroff · 2026-08-16
- User generates 30s 80s arcade-style short film with Dreamina — azed_ai · 2026-08-16
- RTX 5090 turns full manga chapter into animated book in ~3 hours, fully automated and local — Ill-Ant-9489 · 2026-08-16
- Ghost Genesis: Machine Learning Meets Medieval Scripture — misovalko · 2026-08-16