Optimizer Config Prodigy

Prodigy is the optimizer that made a lot of people stop obsessing over learning rate. Its whole pitch is adaptivity: you set the learning rate to 1.0 and it figures out the actual rate on its own as training goes. For the SDXL era that was a gift - finding a good LR by hand is tedious and easy to get wrong - and Prodigy became the community default because of it. OptimizerConfigProdigy is the node that wires Prodigy into a Flux Trainer run with its specific knobs exposed.

A reality check up front: Prodigy's dominance was an SDXL-era thing. On newer architectures the recipes people publish increasingly name AdamW8bit or Adopt-style optimizers instead, and Prodigy notably does not play well with Chroma. It's still a perfectly good choice, and a genuinely convenient one if hand-tuning LR isn't your idea of fun - just know it's no longer the automatic answer it once was.

How it works

The node outputs optimizer_settings (type ARGS) for the Init node. Remember Prodigy's core trick: set the learning_rate on the Init node to 1.0, not a tiny number, and let Prodigy scale from there. That's the part people get wrong.

The inputs worth knowing:

• weight_decay (default 0) - regularization. A small value can help against overfitting.
• decouple (default true) - decoupled weight decay (the "AdamW" style). Leave it on; it's the recommended mode.
• use_bias_correction (default false) - an optional Prodigy stabilizer, occasionally worth trying if early training is jumpy.
• lr_scheduler (default constant) - with an adaptive optimizer a constant schedule is common, since Prodigy is already adjusting the effective rate.
• min_snr_gamma (default 5) - standard loss weighting, leave at 5.

When to reach for it

Use Prodigy when you don't want to hunt for a learning rate and you're on SDXL or Flux (not Chroma). It's the "just let it figure it out" button. If you're following a modern recipe that specifies an exact LR and a plain optimizer, though, there's no need to override it with Prodigy - the auto-tuning is solving a problem that recipe already solved for you.

Installing the pack

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

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

You set a tiny learning rate and it barely trains. The classic Prodigy mistake. Set the Init node's learning_rate to 1.0 - Prodigy interprets that as "scale freely," and a value like 1e-4 hobbles it.

Don't use it on Chroma. Chroma gradient-explodes with the wrong optimizer and Prodigy specifically doesn't work well there; that base wants Adam8bit with very low weight decay and a low fixed LR instead.

It's not magic on new architectures. Prodigy auto-tunes the LR, but it doesn't fix a bad dataset, wrong rank, or too many steps. If results are off, look at those first - the optimizer is rarely the bottleneck.