Solo dev measures anime body proportions in LoRA datasets, finds a negative prompt was shrinking her head
PrudentRespect5648 · reddit · 2026-11-03
A solo dev shares a devlog on AoRA Maker, a local tool for preparing anime character LoRA datasets, obsessed with keeping body proportions stable across poses.
- Measures head units, limb/torso ratios and widths using combined pose, silhouette and face models
- Marks measurements as "not verified" when poses make them unreliable, instead of inventing numbers
- Human stays in the loop: a 15-image blind review showed the auto-tagger missed most declared clothing pieces
- Key find: a precautionary age-related negative prompt was shrinking the character's head, explaining most of the gap between measured and canon proportions — invisible to the eye
The author admits it's unproven that measurement-based image selection improves training, and asks LoRA trainers for validation designs and what breaks first in their character LoRAs. Free devlog on Ko-fi, nothing for sale.
More from Multimodal
- Stunning AI-generated video shows crystals blooming like flowers on an alien world — the-quiet-unreal · 2026-11-03
- 45-second AI ad film "Frenzy in the Air" uses absurd escalation storytelling — LeadingPrestigious52 · 2026-11-03
- Liquid AI Launches LFM3 Model Family; 2B Vision Model Runs Live Camera on iPhone Offline — JosephJacks_ · 2026-11-03
- Creator pairs a Claude Music Video skill with Magnific MCP to produce AI music videos — AIandDesign · 2026-11-03
- Study: multimodal reasoning failures stem from a 'reference gap', not perception — Guizhen Chen · 2026-11-03
- Claude Music Video Plugin generates a full country duet music video — AIandDesign · 2026-11-03