FLUX.1 LoRA dataset requirements
A FLUX.1 LoRA dataset is a folder of images with matching caption text files. Aim for 15 to 25 varied, identity-consistent images at roughly one megapixel, write natural-language captions that start with a trigger word, and keep original aspect ratios rather than cropping everything square. Check the FLUX.1 licence before commercial use.
Published 5 September 2026 · 3 min read
What resolution and aspect ratio should the images be?
Black Forest Labs states that FLUX.1 supports resolutions from 0.1 to 2.0 megapixels and a range of aspect ratios. Most FLUX LoRA trainers expose a base training resolution of 512, 768, or 1024 pixels and bucket images by aspect ratio, so a 3:4 portrait and a 16:9 environmental shot can sit in the same dataset without cropping. Keep source images at or above the training resolution, and do not upscale small images to reach it.
How many images does a FLUX.1 LoRA need?
Fifteen to twenty-five clean, varied images is a comfortable range for one character on FLUX.1. The model is large, 12 billion parameters for FLUX.1 [dev], so each training step costs more than on SDXL, and its T5 text encoder lets captions carry more of the variation. Prefer fewer images that add new angles, framings, and lighting over more images that repeat the same view. The image-count guide has the full reasoning.
How should FLUX.1 captions be written?
FLUX.1 pairs a CLIP encoder with T5-XXL, so it reads natural language well. Write one to three sentences per image: trigger word first, then the variable attributes such as outfit, pose, setting, lighting, and medium. Leave out the fixed identity traits you want the LoRA to learn. Comma-separated tag lists still work but waste the strength of the text encoder.
What file layout do FLUX trainers read?
One PNG or JPEG per image and one text file with the same basename holding the caption. Most FLUX trainers, including ai-toolkit and the FLUX branch of Kohya sd-scripts, read this layout directly from a folder; the Civitai on-site trainer reads it from a zip. LoRA Dataset exports this exact structure.
What should you check before training?
- Every image is the same person. Remove any that drifted.
- Frontal, three-quarter, and profile angles are all present.
- Framing spans close-up to full-body.
- Lighting varies without changing apparent skin tone.
- No second person, watermark, or heavy filter.
- Every caption starts with the same trigger and names only what varies.
What about hardware and the licence?
Training a 12B-parameter model needs more GPU memory than SDXL; trainers offer quantised and low-VRAM modes, and cloud trainers remove the constraint entirely. Separately, FLUX.1 [dev] is released under a non-commercial licence, while FLUX.1 [schnell] is Apache 2.0. If the LoRA will be used commercially, read the licence of the exact base model you train on and generate with.
Frequently asked questions
- Can I reuse my SDXL dataset for a FLUX.1 LoRA?
- The images, yes. Rewrite the captions as natural-language sentences to get the most from the T5 encoder, and check that resolution is around one megapixel.
- Does FLUX.1 need a class noun after the trigger?
- It helps the model know what kind of thing the trigger is. "bunn, a woman wearing..." is a safe pattern; a bare trigger with no noun works less reliably.
- Which trainer should I use for FLUX.1?
- ai-toolkit and Kohya sd-scripts are the common local options; Civitai offers on-site training. All three accept the image-plus-caption folder layout described here.