Customize your input with more control.
Customize your input with more control.
Hint: Drag and drop files from your computer, images from web pages, paste from clipboard (Ctrl/Cmd+V), or provide a URL.
Customize your input with more control.
The cost of training depends on the number of steps. The formula is: 0.0064 * steps. With 1000 steps, your request will cost $6.40.
Fine-tune your training parameters and start right now.
Custom model specialization through LoRA fine-tuning for text-to-image generation. The FLUX.2 [dev] trainer enables you to teach the model your brand's visual language, specific subjects, artistic styles, or specialized rendering requirements through efficient Low-Rank Adaptation training. Train once, generate infinite variations—your custom model deploys instantly to FLUX.2 [dev] LoRA inference endpoints.
Built for: Brand consistency training | Character design | Custom artistic styles | Product-specific rendering | Domain specialization | Style transfer learning
LoRA (Low-Rank Adaptation) fine-tuning specializes FLUX.2 [dev] without the computational cost of full model retraining. By training on your curated dataset, the model learns to generate images that reflect your specific requirements—whether that's maintaining brand visual standards, rendering particular subjects, or applying custom artistic styles.
What you can train:
The quality of your training dataset directly determines model performance. Invest time in collecting and preparing high-quality training data.
Pro tip: Very high-resolution source images (4K+) can be resized during training, which naturally filters out minor imperfections and compression artifacts.
Organize your dataset as a ZIP archive with images and optional caption files:
dataset.zip ├── image_001.png ├── image_001.txt (optional caption) ├── image_002.jpg ├── image_002.txt (optional caption) ├── image_003.png └── ...
Caption files (.txt) should share the same root filename as their corresponding images. If no caption file is provided, the trainer uses the trigger_word parameter as the default caption.
