$5k kale-img-1 fine-tuning contest won with MOSS-Image base, SDXL trails far behind
ostrisai · x · 2026-10-27
A $5,000 bounty asked people to fine-tune kale-img-1, an image model for kids: the host supplied the dataset, participants picked any commercially licensable base model, and entries were judged on recreating 5 dataset images nearly 1:1 plus generating 5 new images in the same style.
Results (same prompt, seed, resolution):
- @ostrisai / 8.10 / MOSS-Image
- @AmandineFlachs / 7.61 / MOSS-Image
- @suhaaspk / 7.20 / MOSS-Image
- @AssuBhadhu / 6.90 / MOSS-Image
- @maxpe2002 / 5.35 / HiDream-O1
- @hasnaindev / 3.51 / SDXL
The top four all used MOSS-Image as the base, clearly outperforming SDXL and others.
More from Multimodal
- Matt Turck generates an Anthropic S-1 explainer video with a single unedited prompt — mattturck · 2026-10-27
- Local text-to-video trailer: Maestro on a 3080Ti plus Premiere Pro Beta's AI Assistant auto-editing — cocktailpeanut · 2026-10-27
- Gen games don't need to be realtime — turn-based interactivity works, devs argue — djcows · 2026-10-27
- Lip sync face-off: Omniflash vs Minimax H3 on The Everlight — Uszko_Visuals · 2026-10-27
- Short film 'Odissey of Aeons' made with MiniMax H3 video generation — Musikalandia · 2026-10-27
- Remiqora Video: open-source local studio turns one idea into 4K chained AI videos — inikolax · 2026-10-27