FluxTrain ModelSelect

Before you can train a Flux LoRA, the trainer needs to know exactly which files to load - and Flux isn't one file. FluxTrainModelSelect is the little node that gathers all of them into one bundle: the transformer (the actual Flux model), the VAE, and both text encoders. Its output is the flux_models input that every Flux Init node demands.

It's not glamorous, but it's where a surprising number of Flux training runs go wrong, because Flux is picky about which version of each file you hand it.

How it works

Four required dropdowns, each picking a file from your ComfyUI models folders:

• transformer - the Flux model itself. For LoRA training this needs to be a normal fp8 or fp16 Flux file.
• vae - and this is the one that bites people: it has to be the non-diffusers ae.safetensors, the single-file VAE from the FLUX.1-dev repo. The diffusers-format VAE folder will not work.
• clip_l - the CLIP-L text encoder.
• t5 - the T5-XXL text encoder.

There's also an optional lora_path string, which lets you point at an existing LoRA to load on top - useful for continuing from or building on a previous adapter. The single output, flux_models (type TRAIN_FLUX_MODELS), plugs straight into InitFluxLoRATraining, InitFluxTraining, or FluxKohyaInferenceSampler.

The gotcha worth repeating

Almost everything that matters about this node is "use the right file." The README is explicit: for LoRA training the models must be the normal fp8 or fp16 versions and the VAE must be the non-diffusers ae.safetensors; for full training you need the fp16 main model specifically. Get these right here and the rest of the training graph tends to just work; get them wrong and you'll hit confusing errors deep inside the Init node that have nothing to do with your actual settings.

Installing the pack

ComfyUI Manager: search ComfyUI Flux Trainer, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-FluxTrainer
pip install -r ComfyUI-FluxTrainer/requirements.txt

Torch 2.4.0+ recommended, plus kijai's ComfyUI-KJNodes for the example workflows.

Common issues

Wrong VAE. The most common one. If training errors out loading the VAE, you almost certainly grabbed the diffusers version. Download the single-file ae.safetensors from the FLUX.1-dev repo and select that.

Your files don't show in the dropdowns. The dropdowns list what's in your ComfyUI model folders. If a file isn't there, drop it into the right subfolder (unet/diffusion_models for the transformer, vae for the VAE, clip/text_encoders for the encoders) and refresh.

fp8 vs fp16 mismatch for full training. LoRA is happy on fp8; full fine-tuning wants the fp16 main model. If you're using InitFluxTraining, make sure the transformer you picked here is the fp16 one.