MiniMax H3 Video LoRA Training Fails: Unable to Capture Identity
LordDarthShader · reddit · 2026-08-11
A developer reports a critical issue when training a LoRA for the MiniMax H3 model using Diffsynth.
- Symptoms: After almost 40 hours of compute, the model picked up clothing and environment but completely failed to capture the subject's identity (face and body), generating random Asian faces.
- Troubleshooting: Even with basic unit-level testing on a single 124-frame video with a very narrow prompt, there is 0 resemblance after 100 steps.
- Comparison: The author previously trained LoRAs for WAN2.2 and LTX using Diffsynth without issues.
The developer suspects a specific bug with MiniMax H3 in Diffsynth and plans to try AI-Toolkit next.
More from Multimodal
- Security Researcher Jeremi Gosney Says Goodbye to Suno AI — nptacek · 2026-08-11
- ACE-Step Workflow: Negative Prompts Are Key to Predictable Background Music — Illustrious_Usual_10 · 2026-08-11
- AI Video Generation Tip: Sync with Audio Using Reference Images and Prompt — techhalla · 2026-08-11
- Agent Workflow Generates 2K Images and Multi-Angle Character Sheets in Minutes — techhalla · 2026-08-11
- Workflow Sharing: Creating AI Music Videos with the H3 Model — techhalla · 2026-08-11
- Local AI Video on RTX 3060 Ti: Generate High-Quality Clips in 20 Minutes with 8GB VRAM — cocktailpeanut · 2026-08-11