Redditor trains Seedream 5.5 style LoRA for MiniMax H3 in 6K steps, works on video too
NebSH83 · reddit · 2026-11-20
Reddit user NebSH83 released an open-source style LoRA that transfers Seedream 5.5's photographic look — lighting, contrast, and film grain — to MiniMax H3 Image.
- Trained in two stages: 2,000 steps on 43 images, then resumed to 6,000 total steps with an expanded 1,134-image dataset (community contributed 150 images)
- Usage: trigger word seedreamstyle, LoRA strength 1.0, LCM 6 steps, CFG 1
- Compared base model, 2K and 6K checkpoints against native 2K Seedream references at 40 MP; also tested on H3 video
Available on Hugging Face and Civitai, with the author soliciting feedback across subjects and lighting conditions.
Related event: Seedream-Style LoRA for MiniMax H3 Image Tops HF Trending(3 posts)→
More from Multimodal
- Ramayana epic recreated as a GTA-style boss battle with AI video and dubbing — umesh_ai · 2026-11-21
- MiniMax H3 video editing: structured prompt swaps summer forest to snowfall seamlessly — smereces · 2026-11-21
- Indie dev open-sources dflow_2b, a 2B text-to-image model trained on a single GPU — quantiler · 2026-11-21
- Ark: an AI-made musical horror tale set in a world where the sun never rises — SoullessGamesDev · 2026-11-21
- AutoCut + FrameAtlas: open-source local-first video editing driven by Claude Code via MCP — moheshmohan · 2026-11-21
- 10 ChatGPT image prompts that generate full brand design kits in 2 minutes — anthara_ai · 2026-11-21