Training multi-character LoRAs by having Claude Opus write the training code
Amazing_Painter_7692 · reddit · 2026-10-29
The author trained a multi-character LoRA on Krea2 Turbo BBOX by simply feeding group images of Smiling Friends characters to Claude Opus 5.5 with one instruction: draw bboxes per the prompting guide, write LoRA training code, and train. Setup: AdamW, LR=1e-4, attention layers plus text fusion layer trained, images labeled per the public prompting guide. Before/after comparison shows markedly improved character consistency.
Related event: Claude Opus Automates Multi-Character LoRA Training With a Single Prompt(2 posts)→
More from coding & agent
- Claude Artifacts creator launches Spacesheep, a YC F26 shared workspace for humans and agents — ycombinator · 2026-10-29
- TradeOS AI Hits 1,600 Installs in a Week: Claude/GPT Can Now Analyze XAUUSD Charts — DangerousFile467 · 2026-10-29
- OpenKatana: An Open-Source AI Video Editor That Lets Claude Drive ComfyUI — rk0neki · 2026-10-29
- Indie dev turns AI-generated walking videos into natural game sprite animations — yshan2u · 2026-10-29
- gemini-cli PR fixes flaky test by mocking realpathSync in orphan PTY slave fd case — AbbasFullstack · 2026-10-29
- He vibecoded an AI attractiveness rater validated on real data; it scored him 5.1 — AaronBergman18 · 2026-10-29