How to train a FLUX.3 Image Klein LoRA

Which trainers support FLUX.3 Image Klein (BFL), how much VRAM you need, and settings that work. Every row links to its source.

Trainer support

Trainer	Status
ai-toolkit	supported (day-0, PR #1078 merged Oct 15; fixes Oct 15)
OneTrainer	unknown
musubi-tuner	unknown
kohya sd-scripts	unknown
SimpleTuner	supported (Base + distilled klein recipes, 24-144 GB)
diffusers	inference only (v0.42.0, Oct 23; Base and distilled via ModularPipeline, FP8 and NVFP4 checkpoints)

SimpleTuner recipes: Base from 24 GB (512px, batch 3) up to 144 GB; distilled klein at 24 GB (768px, batch 1) or 80 GB (1024px, batch 4)

Open weights, gated on Hugging Face: click 'Agree and access' on the model page or downloads fail with 403
Mixed early reception; watch for anatomy issues
Base model sampling: ~50 steps, CFG 3.5
LoRA strength ~1.3 at inference
Character LoRAs come out weak at the default rank 16/32; go higher
SimpleTuner recipes: rank-32 LoRA, LR 1e-4, REPA; distilled klein auto-loads a published training-assistant LoRA (strength 1.0 in training, 0.0 at validation; disable_assistant_lora to turn off)
ai-toolkit edit training with several control images used only one until a fix on Oct 17; update before training edit LoRAs
A FLUX.2 dataset and recipe carried over unchanged doesn't work well; long upsampled prompts with JSON bounding boxes at ~2 MP sample better
diffusers v0.42.0 (Oct 23) supports FLUX.3 Image Klein Base and distilled through ModularPipeline only (ModularPipeline.from_pretrained, then load_components), with the prompt upsampler and BFL FP8 and NVFP4 checkpoints. Needs Python 3.11+
Train and generate at 1.5k resolution or higher rather than 1k; the VAE holds it back at 1k
Fast previews of a klein base LoRA: the FLUX.3 klein turbo PDD LoRA runs base at 8 or 4 steps with CFG distilled in (text to image; image to image loses quality)

Gated Hugging Face models: click 'Agree and access' on each model page, or you'll get 403 errors
Steps scale with dataset size (steps = epochs x images at batch 1)
A safe start: LoKr with auto rank (~64), batch/grad accumulation 4
Leave out of your captions whatever you want burned into the LoRA; caption what should stay changeable
ai-toolkit dataset folder weights let you train the same images with long, medium and short captions in separate folders
ai-toolkit supports validation loss, which shows when a run stops improving
In ai-toolkit, Cache text embeddings works for basically every model and frees the text encoder from VRAM and RAM
Prefer int8 convrot over qfloat8 quantization in ai-toolkit: better quality at the same memory
Gradient accumulation and batch size don't scale the learning rate in ai-toolkit; a bigger batch needs about 1.5-2.5x the steps

Weekly newsletter: new models, trainer releases, cheapest GPUs. Subscribe

Data updated 2026-10-20. RunPod and Vast.ai links are referral links. unknown = no data yet. · Raw data (JSON) · llms.txt