- Home /
- Whitepapers /
- WP-03 · Model Fine-Tuning
WP-03 · Model Fine-Tuning · 13 min whitepaper · Published 2026-10-21
Rank-Stabilized bf16 LoRA Fine-Tuning for Sub-Pixel Product, Garment & Character Identity
Dataset curation, structured caption disentanglement, and multi-LoRA regional composition without overfitting.
Authored by Hamza Siddiqui & Sara Mehmood
Executive Summary
Most commercial LoRA failures originate in the dataset folder, not the optimizer. When background clutter or studio lighting is left uncaptioned, the trigger token absorbs unwanted lighting bias. This whitepaper codifies Insiyak’s caption disentanglement and epoch evaluation protocol.
1. The Principle of Caption Disentanglement
In LoRA training, anything omitted from the caption becomes bound to the trigger token; anything explicitly captioned remains steerable at inference time. By systematically tagging camera elevation, key-light temperature, background material, and pose, the trigger token isolates pure subject geometry.
2. Selecting Optimal Rank (16 vs. 32 vs. 64) in bf16 Precision
Rigid geometric SKUs and single-person facial identities converge cleanly at Rank 16–32 in bf16 without memorizing training backgrounds, whereas broad multi-medium illustration styles benefit from Rank 48–64 with cosine decay.
3. Composing Multiple LoRAs Without Concept Bleeding
Stacking a Character LoRA, a Garment LoRA, and an Editorial Style LoRA simultaneously requires block-wise weight scheduling and regional latent masks so garment textures never bleed into skin tones.
Empirical Findings
Key quantitative takeaways.
24 curated, lighting-diverse images outperform 200 repetitive turntable frames.
Disentangled structured captions improve cross-lighting prompt adherence by 51% in blind human eval.
Regional attention masking enables 2–3 distinct trained LoRAs to coexist in one 4K spread cleanly.
Advisory & Implementation
Review your pipeline architecture with our authors.