alvdansen/linnea-qwen-image

Linnea — Qwen-Image

A character LoRA for Qwen-Image trained on 27 illustrated reference images of an original character — Linnea, mint-green hair, a recurring motif in Alvdansen character work.

Published alongside its straight-baseline twin as the small-dataset paired comparison in Forgetting on Purpose: Generalization as the Quality Criterion for Small-Dataset LoRA Fine-Tuning — Alvdansen Labs, May 2026. Read the paper · Source on GitHub.

This is the publication-grade checkpoint at step 35,000, selected from a dense consolidation-phase probe.

Usage

Trigger word: linnea

Compose prompts naturally around the character — portrait, full-body, scenes, expressions. Adding suffixes like , illustrated style reinforces the trained illustration register.

Recommended Inference Settings

Sampler: euler
Scheduler: simple
CFG: 3.5
Steps: 45 (30–60 works well)
LoRA strength: 0.8–1.0

Training Details

• Base model: Qwen-Image (FP8 quantized, text encoder FP8)
• Training steps: 35,000 (selected publication checkpoint; training ran to step 35,750)
• Schedule: 4-phase chained — three 9-image subsets trained sequentially, then the full combined 27-image dataset reintroduced for consolidation
• Rank/Alpha: 42/42
• Learning rate: 5e-5
• Optimizer: AdamW 8-bit
• Caption dropout: 0.25
• EMA: enabled (decay 0.99)
• Noise scheduler: flowmatch
• Precision: bf16 with qfloat8 quantization
• Dataset: 27 illustrated reference images, 3 disjoint subsets of 9 each
• Trainer: ai-toolkit by Ostris
• Hardware: NVIDIA RTX 6000 Ada (A6000, 48 GB VRAM)
• Wall-clock: ~24 hours

Full configuration in Appendix A of the paper.