SimpleTuner/Qwen-Image-2.1-LoRA-experiments

Qwen Image 2.1 LoRA experiments

Built with Qwen. Experiments in training Qwen Image 2.1 LoRAs with SimpleTuner, covering training assistants, synthetic regularisation, resolution, REPA, and longer training. Chapters include sample comparisons, observations, checkpoints, and reproduction recipes.

Experimental ablations to find ideal settings

Follow the experiments from baseline training through the choices behind the successful runs.

1. Baseline concept training — How ordinary training changes the character and unrelated subjects.
2. Training the assistant: v1 and v2 — How synthetic and mixed-data assistants change the base model.
3. Concept training with a frozen assistant — Character learning, prompt wording, and 512px versus 1024px training.
4. Synthetic regularisation — Comparing base-prediction matching with assistant-based training.
5. Combining an assistant with regularisation — Training with an assisted parent target at two resolutions.
6. Multi-scale training and REPA — Comparing resolution mixtures, feature alignment, and flow shift.
7. Learning-rate experiments — The degradation observed in the 3e-4 run.
8. Longer training: regression and recovery — What checkpoint progression teaches us about when to stop.

Successful training runs

• Domokun: multi-scale, REPA, and auto shift — The preferred 10,000-update checkpoint, complete training settings, and all five prompts across seven checkpoints.
• Photo-aesthetics: 50,000 updates — Extended 512px training with assistant v2, plus a 1024px-trained reference.
• Photo-aesthetics v2: multi-scale, REPA, and auto shift — The 50,000-update follow-up, with checkpoints at 10,000-update intervals.

Reproduction and reference

Shared training and evaluation settings · Using the weights and recipes · Data and license

Limitations

The comparisons are qualitative and use one training seed per variant. Prompt, guidance, and evaluation settings are recorded with each grid. Historical comparisons differ in more than one training variable; findings apply to the shown runs and samples.