Redder trains LoRAs for MiniMax H3 with ai-toolkit, shares first local text-to-video clips
renderartist · reddit · 2026-10-24
A Reddit user shared first results from training LoRAs for MiniMax H3 using ai-toolkit. With many model variants and configurations to sort through, it took him about two days to get started, after which he produced his first clips fully locally via text-to-video — no first frame or reference images. He was impressed by how well the model picked up the style characteristics he was aiming for. The attached image is fully black and provides no additional visual information. A useful hands-on signal for anyone experimenting with local fine-tuning of video generation models.
More from Multimodal
- China's first fully AI-generated theatrical film took 2 years and 1.2M images — lmoroney · 2026-10-24
- Simplified Matching Distillation: few-step diffusion without adversarial training, validated up to 33B — ArashVahdat · 2026-10-24
- Kling 4 hands-on: near-perfect 4-character video swap on first try, faster and cheaper than Seedance 2.5 — venturetwins · 2026-10-24
- An AI music video for everyone still building something at 3am — Objective_Task1588 · 2026-10-24
- Redditor Tests 100+ h3 Turbo Video Configs, Builds Blind Ranking Site — TypeItRight · 2026-10-24
- Dev Vibe-Codes MOSS Image Studio, a Friendlier Browser Frontend for ComfyUI — Roorem10 · 2026-10-24