FLUX LoRA Training With 256 Images: Full Workflow & Conclusions
FLUX LoRA training experiments with 256 images on 8x A6000 and 1x L40S GPUs: 80 epochs worked best, and doubling the learning rate for 8 GPUs overtrained.

Highlights
Three trainings with Kohya configs, full workflow and conclusions: keep LR 0.0001 on 8 GPUs, avoid repeating clothing and backgrounds, use sharp images with varied expressions and distances. The model learned fine details, expressions and distant shots; grids are on Hugging Face.
What it does
Three FLUX LoRA trainings on a 256-image dataset of one person, made to show that a dataset with expressions and different perspectives lets the model generate them well. The configs come from the main Kohya FLUX LoRA training post.
What you get
- Test 1: 8x A6000 with the 4x_GPU_Rank_1_FAST_Lower_Quality config and a doubled learning rate of 0.0002, which overtrained after 40 epochs
- Test 2: a single L40S with the Rank_1_T5_XXL_39200MB config up to 40 epochs, which was undertrained
- Test 3: 8x A6000 with the 4x_GPU_Rank_1_SLOW_Better_Quality learning rate plus T5 XXL training
- Comparison grids for each test and the best result: 80 epochs, which is 2,560 steps on 8 GPUs
Conclusions
- Keep the learning rate at 0.0001 when you move from 4 to 8 GPUs; doubling it overtrains quickly
- 256 images with repeating clothing and backgrounds overfit, so collect varied expressions, perspectives, distances and clothing
- Use images with the best possible sharpness, focus and lighting
- Only 'ohwx man' was used as the caption, since captioning reduced likeness
- The model learned fine details, expressions and distant shots
Related post
Related

Qwen Image LoRA Training: Stage 1 Results and Pre-Made Configs
Qwen Image LoRA results from more than 35 Stage 1 trainings with SECourses Musubi Tuner; the published pre-made configs train on GPUs with as low as 6 GB.

FLUX 2 vs FLUX SRPO & New Kohya SS GUI Premium FLUX Training App
FLUX 2 Pro compared with a trained FLUX SRPO model, plus the new Kohya trainer V35 that trains FLUX on GPUs with as little as 6 GB VRAM.

FLUX LoRA vs Fine Tuning / DreamBooth Comparison: Batch Size 1 vs 7
FLUX LoRA and Fine Tuning / DreamBooth compared with 15 and 256 training images at batch size 1 and 7, for both realism and stylization.