Init Flux Training

Note the missing word: this is InitFluxTraining, not LoRA Training. This node does a full fine-tune of the Flux model - the DreamBooth path - where you're updating the actual model weights instead of training a small adapter on top. It's the heavyweight cousin of InitFluxLoRATraining, and it's the one you reach for when a LoRA isn't cutting it and you want a full checkpoint out the other end.

Full fine-tuning consistently produces better quality than a LoRA - that's the settled community verdict - but it's far more demanding on hardware and it produces multi-gigabyte checkpoints instead of a tidy few-megabyte adapter. The usual advice holds: LoRA for personal and creative projects, full fine-tune for a base model you intend to distribute widely. Kijai's README flags this path as "untested" relative to the LoRA nodes, so go in expecting rough edges.

How it works

Same shape as the rest of the pack. Required inputs are flux_models (from FluxTrainModelSelect), a dataset JSON, and optimizer_settings. It builds the trainer and outputs network_trainer (NETWORKTRAINER) for a FluxTrainLoop node to step through, plus epochs_count and raw args. The difference from the LoRA node is what's not here - there's no network_dim/network_alpha, because you're not building a low-rank network, you're training the model itself.

The inputs that matter

• learning_rate (default 4e-6) - much lower than the LoRA default, and correctly so. Full fine-tunes want a gentler learning rate because every weight is in play.
• blocks_to_swap - this is the load-bearing knob for full Flux fine-tuning. Block swapping offloads transformer blocks to system RAM, and it's what brought Flux DreamBooth down to famously low VRAM figures. Expect to need 32GB+ of system RAM to lean on it.
• cpu_offload_checkpointing (default true) and fp8_base - more memory relief. Full fine-tuning is a VRAM fight and these are your tools.
• optimizer_fusing (fused_backward_pass or blockwise_fused_optimizers) - a memory optimization specific to the full-training path that fuses the optimizer step into the backward pass. Leave it on the default unless you have a reason.
• guidance_scale, timestep_sampling, discrete_flow_shift - the Flux flow-matching controls, shared with the LoRA node; the defaults are the starting point.

Installing the pack

ComfyUI Manager: search ComfyUI Flux Trainer, install, restart. 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. Critically, for full model training the README says you need the fp16 version of the main model - not the fp8 file you might use for LoRA work. And your VAE has to be the non-diffusers ae.safetensors.

Common issues

You almost certainly want a LoRA instead. Be honest about why you're here. Full fine-tuning eats VRAM, time, and disk, and for a character or style the community's near-universal answer is that a well-trained LoRA gets you most of the way at a fraction of the cost. Reach for this node when you're building a base model to hand out, not to teach Flux one face.

Wrong model precision. If you loaded the fp8 model expecting full training to work, it won't behave - full fine-tuning wants the fp16 main model per the README.

It's the untested path. Kijai's own words. The LoRA nodes get the attention; this one is more likely to surprise you. Do a short validation run and watch VRAM before you commit to a long job.