Flux Train Resume

Training got interrupted - a crash, an OOM, you closed the tab, the power blinked - and you'd rather not throw away the two hours of GPU time you already spent. That's what this node is for. FluxTrainResume points a new run at a saved training state and continues from there, optimizer momentum and all, instead of starting from scratch.

The catch, and it's the whole story: you can only resume if you saved state in the first place. Resume reads a file that has to already exist.

How it works

There's an important distinction between a saved LoRA and a saved state. The LoRA is just the adapter weights - great for using the model, useless for continuing training, because it doesn't remember where the optimizer was. A saved state is the full training snapshot: weights plus optimizer momentum, step counter, the works. FluxTrainer writes that when you enable save_state on the FluxTrainSave node. FluxTrainResume is the other half of that pair - it loads the state back so the run picks up mid-stride rather than cold.

It produces a resume_args bundle (type ARGS) that you feed into the training-init node, alongside the other optional config bundles. Wire it in, point it at your state file, and the run resumes.

The inputs that matter

Only two, and both are simple:

• load_state_path - the path to the saved state directory/file from your earlier run. This is a multiline string field; paste the real path. If you enabled copy_to_comfy_lora_folder and save_state earlier, the state lives wherever FluxTrainer wrote the checkpoint.
• skip_until_initial_step - off by default. When on, it fast-forwards past the steps already completed instead of re-running them, so your data ordering and schedule line up with where you actually left off. Turn it on for a true continuation.

Output is resume_args.

Installing ComfyUI-FluxTrainer

ComfyUI Manager: search "ComfyUI Flux Trainer", install, restart. Manual:

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

Restart ComfyUI. Torch 2.4.0 or newer recommended. The example workflow pulls in ComfyUI-KJNodes.

Common issues

No saved state, no resume. This is the mistake, and it's unrecoverable after the fact. If you didn't set save_state to true on FluxTrainSave during the original run, there's nothing for this node to load - you'll have LoRA checkpoints but no way to continue, and you start over. The lesson is prospective: if a run is long or you think you might extend it, enable state saving before it starts. State files are large, which is exactly why saving them is off by default and why people forget.

Path problems. The other failure mode is a wrong or moved load_state_path. Point it at the actual saved state, use the full path if the relative one doesn't resolve, and don't rename or relocate the folder between saving and resuming.

Config drift. Resume works best when the rest of your graph matches the original run - same dataset config, same model, same network settings. Changing the recipe and then resuming is asking for weirdness. If you want different settings, that's a new run, not a resume. Treat this node as "continue the exact thing I was doing," and it's reliable; treat it as "restart with edits," and you'll fight it.