Question about lora fine tune qwen-image-edit

Research
1
  1. What’s the suitable dataset size and epochs?
  2. Can I train several tasks by training one time? Distinguish them by dataset and captions.
1 Like
2

maybe

  1. Dataset size + epochs (for Qwen-Image-Edit LoRA)

    • Practical starting range (per task):

      • 20–50 pairs for a narrow thing (one product / character / angle set).
      • 80–200 pairs if you want a more general edit style that works on many scenes.

    • Training length (rough):

      • Aim for about 1k–5k training steps total. With 40–100 images and batch size 2–4, this usually means 100–200 epochs.

    • Keep images and captions high-quality; more low-quality images usually makes the LoRA worse, not better.

    Good Qwen-specific tutorial that uses these kinds of numbers:

  2. Several tasks in one training run?

    • Yes, you can train one LoRA on several tasks at once. This is common in “multi-concept” LoRA training.

    • How to do it in practice:

      • Mix all examples into one dataset.
      • Use different trigger words / phrases in the captions to distinguish tasks (e.g., <angle_style>, <skin_fix>, <logo_style>).
      • Keep each task roughly balanced in number of samples (don’t let one task have 5× more images than the others).

    • Trade-offs:

      • One multi-task LoRA is convenient but can cause interference between tasks.
      • For very different tasks, many people prefer separate LoRAs per task and then combine them at inference time (load multiple LoRAs with different strengths).
3

Oh wow. Thats so little data actually! Has any1 tried w only 40-100 samples? (i ve only trained LLMs, have no experience w image gen)

1 Like
4

e.g. https://note.com/tori29umai/n/n256f30d51669 (in Japanese)

1 Like
5

hello everyone..!