FLUX.2 is now live!
DocsBlogPricingEnterpriseCareersResearch Grants
Log-inSign-up

Flux 2 Trainer Training

fal-ai/flux-2-trainer/edit
Fine-tune FLUX.2 [dev] from Black Forest Labs with custom datasets. Create specialized LoRA adaptations for specific editing tasks.
Training
Commercial use
Schema
LLMs
PlaygroundAPI

Input

Additional Settings

Customize your input with more control.

The cost of training depends on the number of steps, reference images. The formula is: 0.009 * steps * reference_multiplier. The reference multiplier for 1, 2, 3 and 4 images is 2.11, 3.44, 5.09, and 6.95 respectively. With 1 reference image and 1000 steps, your request will cost $18.99.

Training history

See all
Note: these are the most recent training requests. For the full history, check the requests tab.

FLUX.2 [dev] Edit Trainer

Fine-tune FLUX.2 [dev] with LoRA to teach it your custom image-to-image editing behavior. Train once on before/after pairs, then apply consistent, specialized edits forever — no full model retraining needed. A small LoRA (a few MB) captures exactly how you want images transformed, giving you precise control that generic models or prompting alone can’t achieve.

Perfect for:

  • Brand-consistent edits: Teach adjustments that maintain visual brand guidelines across transformations
  • Product-specific processing: Specialized understanding of how to modify particular product categories or materials
  • Style transfers: Custom style transfers that generic models struggle with
  • Technical/domain-specific corrections: Industry-specific editing patterns for specialized domains
  • Automated editing pipelines: Consistent editing behaviors that match your team's established patterns
  • Anything you can show with examples!

What you need

  • 15–50 high-quality before/after pairs (more = better specialization)
  • Resolution: ≥ 1024×1024 (higher is better)
  • Clear, consistent transformation between start → end
  • Clean images, no compression artifacts

File naming (strict!)

XXX_start.png/jpg  → original image (unedited)
XXX_start2.png/jpg → optional extra reference (up to 4 total)
XXX_start3.png/jpg
XXX_start4.png/jpg
XXX_end.png/jpg    → desired edited result
XXX.txt            → optional instruction prompt

Example structure with a single reference image:

editing_dataset.zip
├── 001_start.jpg
├── 001_end.jpg
├── 001.txt          # e.g. "apply brand color grading and remove background"
├── 002_start.png
├── 002_end.png
├── 002.txt
└── ...

Example structure with 3 reference images:

editing_dataset.zip
├── 001_start.jpg
├── 001_start2.jpg
├── 001_start3.jpg
├── 001_end.jpg
├── 001.txt          # e.g. "apply brand color grading and remove background"
├── 002_start.png
├── 002_start2.png
├── 002_start3.png
├── 002_end.png
├── 002.txt
└── ...

`.txt` files are optional but strongly recommended for best results.

If you prefer one caption for everything, just set a Default Caption instead.

That’s it — zip the folder, upload, train a LoRA and get unlimited custom editing intelligence tailored to your exact needs.