[Python] How do I properly train or fine-tune flux.1-dev on a custom emoji dataset? [closed]

I’m trying to train black-forest-labs/FLUX.1-dev on my own dataset of about 25,000 emoji-style images. Each image (512×512 PNG) has a short caption, and I'm trying to generate new emoji-like images from text prompts.

dataset/
├── images/
│ └── image_1.png
├── captions/
│ └── caption_1.txt
└── train.json # [{"image": "images/image_1.png", "caption": "captions/caption_1.txt"}, ...]


I'm not looking to debug anything specific — I just want clear guidance or an example on how to train this model properly. Specifically:

1. 
   How should FLUX.1 be fine-tuned?
   
   • Should I freeze the VAE or text encoder?
   • Should I use LoRA, full fine-tuning, or DreamBooth?
2. 
   Is my dataset format okay?
   
   • Are .txt captions + image pairs + a train.json enough?
3. 
   Do you have an example training script or notebook?
   
   • I couldn’t find official examples specific to FLUX.1.
4. 
   What’s a good config?
   
   • Batch size, learning rate, number of epochs for decent results?

Any kind of walkthrough, training script, or working setup would be helpful.
​

Continue reading...