MOSS-Image SFT Domokun, 512px

Part of the MOSS-Image LoRA experiments.

Training status: completed, 10,000 / 10,000 observed updates. Captions: native JSON captions shared between both flavours. Each run starts fresh from its released checkpoint, without an assistant LoRA. Resolution is 512px equivalent image area with aspect ratios preserved.

Rank-32 BF16 LoRA, batch 4, learning rate 0.0001, seed 42, REPA, one H100 80 GB. Validation uses 50 sampling steps, CFG 3, seed 42, and four prompts every 500 updates.

Recipe

The training config, dataset config, prompts, and dataset manifest record the run. Publishing is handled by an external checkpoint collector; the training config retains the in-process Hub flags used by training. The dataset is RareConcepts/Domokun, revision 2e4831a72f2a701c080e0046c8f8ec41cb1ccbcf, containing 29 images. Dataset files are not redistributed here.

Checkpoints

500, 1,000, 1,500, 2,000, 2,500, 3,000, 3,500, 4,000, 4,500, 5,000, 5,500, 6,000, 6,500, 7,000, 7,500, 8,000, 8,500, 9,000, 9,500, 10,000

Latest published adapter: update 10,000.