Training Style LoRAs for Minimax h3 Yields Underwhelming Results
VividCustard5884 · reddit · 2026-08-13
A developer shared bottlenecks encountered while training style LoRAs for the Minimax h3 video model using ai-toolkit.
Despite extensive experience with various datasets, the training results on Minimax h3 were filled with artifacts and suffered from abnormally low brightness, making them barely usable. The author noted minimal community exploration into style LoRAs for this model and invited discussion, highlighting the potential of successfully training retro-game style video LoRAs.
More from Multimodal
- Flux 3 Klien first impressions: occasional body horror like SD3, otherwise solid — rionix88 · 2026-10-16
- XGRIDS builds world's largest 3DGS town: 23.5 km² and 6.2 billion Gaussian Splats — bilawalsidhu · 2026-10-16
- IMMERSION: AI film compilation links 12 shorts with one continuous camera move — AI_Andrew · 2026-10-16
- A song with vocals, lyrics and 3D head animation created entirely by Opus 5.5 — repligate · 2026-10-16
- Alibaba open-sources CommerceGen-Studio for commercial-grade e-commerce images — alibabagroup · 2026-10-16
- EVA-Cap: Fudan and OPPO align video captioning with atomic audiovisual events via GRPO — Fudan-University · 2026-10-16