alvdansen/embe-qwen-image

Embe — Qwen-Image

A character LoRA for Qwen-Image trained on 28 illustrated reference images of Embe — a recurring antlered character in Alvdansen's drawing work.

Trained with the chained schedule from 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 release ships the last 10 checkpoints from the 4-phase chained run (final at step 6,776). The main embe-qwen-image.safetensors is the consolidation-phase endpoint; intermediate checkpoints are in checkpoints/ for anyone who wants to compare points along the schedule.

Usage

Trigger word: embe

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

Checkpoints

File	Step	Notes
embe-qwen-image.safetensors	6,776	Endpoint of Phase 4 (combined consolidation). Default pick.
checkpoints/embe-qwen-image_step6750.safetensors	6,750	Last 250-step save before endpoint.
checkpoints/embe-qwen-image_step6500.safetensors	6,500
checkpoints/embe-qwen-image_step6250.safetensors	6,250
checkpoints/embe-qwen-image_step6000.safetensors	6,000
checkpoints/embe-qwen-image_step5750.safetensors	5,750
checkpoints/embe-qwen-image_step5500.safetensors	5,500
checkpoints/embe-qwen-image_step5250.safetensors	5,250
checkpoints/embe-qwen-image_step5000.safetensors	5,000
checkpoints/embe-qwen-image_step4750.safetensors	4,750	Start of the consolidation phase region.

Training Details

• Base model: Qwen-Image (FP8 quantized, text encoder FP8)
• Total training steps: 6,776 (4-phase chained schedule)
• Schedule: three 9- or 10-image disjoint subsets trained sequentially (Phases 1–3), then the full 28-image dataset reintroduced for Phase 4 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: 28 illustrated reference images
• Trainer: ai-toolkit by Ostris
• Hardware: NVIDIA RTX 6000 Ada (A6000, 48 GB VRAM)

Captions generated with klippbok's character-mode template (action and setting only, no character appearance description — the LoRA learns the visual identity from the training pixels).

Citation

@article{carlson2026forgetting,
  title   = {Forgetting on Purpose: Generalization as the Quality Criterion for Small-Dataset LoRA Fine-Tuning},
  author  = {Carlson, Minta and Bielec, Timothy},
  year    = {2026},
  month   = {May},
  journal = {Alvdansen Labs},
  url     = {https://alvdansen.github.io/forgetting-on-purpose/}
}