Train Multi-Character LoRAs With One Claude Prompt Using Bbox Cues
Amazing_Painter_7692 · reddit · 2026-10-29
A Reddit user demonstrated an ultra-simple workflow for training multi-character LoRAs with Claude Opus 5.5: given a group photo from Smiling Friends, they simply said "draw boxes following the bbox prompt guide, write the training code, then train the LoRA," and Claude handled annotation and training automatically.
- Training details: AdamW optimizer, learning rate 1e-4, training all attention layers plus text fusion layers
- Bbox prompting: each character gets a coordinate-boxed description in the prompt (e.g., ac:1 Pim, ac:2 Charlie); the included three-way comparison shows steady improvement from no LoRA → LoRA → LoRA+bbox in multi-character consistency
- In the screenshots, character colors remain muddled with LoRA alone, but after adding bbox prompts all four characters' features map precisely
Related event: Claude Opus Automates Multi-Character LoRA Training With a Single Prompt(2 posts)→
More from coding & agent
- AI subagent claims 'work complete', lead agent's audit discipline steals the show — teortaxesTex · 2026-10-29
- He vibecoded an AI attractiveness rater validated on real data; it scored him 5.1 — AaronBergman18 · 2026-10-29
- DeepMind opens Antigravity Science Program, pairing Gemini agents with AlphaFold for research — DeryaTR_ · 2026-10-29
- How one Redditor squeezed Raon-1-mini from 50 to 232 tok/s on a 16GB RTX 5070 Ti — McNiiby · 2026-10-29
- Refactoring one SKILL.md into a Kanban agent graph: same output, 3x tokens, 4x time — granawkins · 2026-10-29
- 10-person Spawn spends 30x payroll on AI tokens, ships 77,000 changes a month — mattshumer_ · 2026-10-29