Kromoss MOSS finetune tested: reuses base LoRAs, plus a rank-32 LoRA training config
Any_Tea_3499 · reddit · 2026-11-22
The author shares hands-on results with Kromoss, Lodestone's finetune of the MOSS image model: it works well and can directly reuse character LoRAs trained on base MOSS without losing likeness, with BF16 and INT8 performing equally.
Workflow details
- Adapts the Er-SDE sampler for better image quality, with a low-strength turbo LoRA at 15-20 steps
- Updated LoRA training config: rank-32 LoRA for small datasets; LoKr 4 still recommended for datasets over 100 images (rank 32 struggles with complex faces)
- Links to Kromoss, the SFT turbo LoRA, Canon/Lenovo style LoRAs on Hugging Face, plus the full workflow and training config downloads
More from Multimodal
- AI video made with Seedance 2.5 on Higgsfield hailed as a masterpiece — SimplyAnnisa · 2026-11-22
- AI image meme "Your Timetraveler": glowing forest tunnel with lone figure goes viral — LudovicCreator · 2026-11-22
- New ComfyUI node runs prompts sequentially from a spreadsheet with wildcard support — ForesterAI · 2026-11-22
- Dev ships ComfyUI node that runs prompts sequentially from a spreadsheet with wildcard support — ForesterAI · 2026-11-22
- Suno's new music video tool hooks into Grok Bot via MCP for full AI-directed creation — tetsuoai · 2026-11-22
- fal Agent lands in Photoshop, demoed to 1,000+ at Adobe MAX — adamho · 2026-11-22