---
tags:
- flux
- text-encoder
- lora
- fine-tuned
---

# FLUX.2 Dev Text Encoder — Fine-Tuned
Mistral 30-layer TE fine-tuned with LoRA (rank 32) on JoyCap Beta captions.
Includes merged safetensors (drop-in replacement) and LoRA checkpoint.

## Files
-  — merged model (35.6 GB)
-  — LoRA adapter checkpoint (529 MB)
-  — training script
-  — training log

## Training
- 3 epochs, 2,562 steps on A100 SXM4 80GB
- Final train loss: 0.78, eval loss: 0.90
- 13,956 JoyCap Beta captions