FLUX LoRA dimensions and dropout

last edited

Advanced FLUX LoRA training involves choices about network dimensions, dropout, licensing, overfitting, and validation. Everyday recipes appear at Flux LoRA Training - Community Techniques. Trainer quirks appear at Kohya Trainer Edge Cases.

Licensing and tooling

Adapters generally inherit the base model's license. The FLUX.1 dev license file restricts weights to non-commercial use. Generated outputs can be used commercially. The same license covers six variants: kontext, fill, canny, depth, redux, and krea.

FLUX.2 klein arrived in January 2026. It has one permissive variant and one non-commercial variant. The 4B model is royalty-free for paid apps.

The developer announced the FLUX 3 family in July 2026. A permissive FLUX 3 dev model is planned for later that year. Training Tool Comparisons covers alternatives. Deployment context appears in Flux Production Usage.

Network dimensions and rank

Common defaults set linear and linear_alpha to 32. Conv and conv_alpha also default to 32. An improved configuration used linear 128 with linear_alpha 64.

That configuration set conv to 64. Its conv_alpha was 32. Higher dimensions increase adapter capacity and training cost. The LoRA research paper describes rank selection. ComfyUI Ecosystem covers training software. Broader context appears in Generative image and video AI research areas.

Dropout, data, and validation

The T-LoRA paper on arXiv, accepted at AAAI 2026, reports that higher diffusion timesteps overfit more than lower ones. The authors publish the T-LoRA official code repository. Reducing rank or expanding the dataset are usual first responses to overfitting and underfitting.

An uncaptioned recurring character in a style dataset can make the LoRA learn that character instead of the intended style. Validation images should keep prompts and inference settings consistent across checkpoints. For inference combinations, consult LoRA Stacking Techniques.

Sources