Moss LoRA training in AI-Toolkit: default rank 32 looks bad, rank 8 lacks detail
Lucaspittol · reddit · 2026-10-19
A Redditor reports trouble training LoRAs for the Moss model in AI-Toolkit: the default rank 32 produces poor results, while rank 8 looks better but lacks detail. Datasets that worked well for Chroma and other models don't transfer well to Moss. They're asking the community for prior experience and dataset configurations that work for Moss — a noteworthy practical data point for anyone fine-tuning image models.
More from Multimodal
- Fish Audio open-sources 35.9B MoE TTS Drama 3, tops leaderboard while decoding faster than its 4.6B predecessor — kastnerkyle · 2026-10-19
- Runway Labs launches as frontier research design studio for video and world models — jongranskog · 2026-10-19
- earshot v0.2: open-source 'ears' that let AI coding agents hear, judge and synthesize sound — teropa · 2026-10-19
- inclusionAI open-sources Ming-3.0-audio-tiny: 9.3B MoE audio-language model (MIT) — jacek2023 · 2026-10-19
- Fast Hours hits 100K YouTube subs with cracked AI creative workflows episode — Ror_Fly · 2026-10-19
- Stacking 4 LoRAs on MiniMax H3 for a cinematic dark sci-fi look — pixellegolas · 2026-10-19