TryOnOff Model Loader [ComfyUI-Flux-TryOff]

Honest answer up front: this node and TryOff Model Loader are the same thing. Same Python class, same inputs, same outputs, same behavior - the pack registers the identical TryOnOffModelNode class under two names ("TryOnOff Model Loader" and "TryOff Model Loader"), so whichever you drag in, you're getting the exact same loader. The only difference is the label on the box.

What it does

Loads one of three Hugging Face fine-tunes of the Flux transformer - xiaozaa/cat-tryoff-flux (try-off), xiaozaa/catvton-flux-beta and xiaozaa/catvton-flux-alpha (try-on) - as a bf16 diffusers FluxTransformer2DModel, downloading it into ComfyUI/models/checkpoints on first use. Inputs:

• model_name - the three-option picker above.
• device - cuda or cpu. Use cuda.
• transformers_config (optional) - from the TryOff Quantizer, for 8-bit/4-bit loading.

Output: a single MODEL (the transformer), which feeds the FluxFill Pipeline Loader or the legacy FluxFill Model Loader. If you wire the quantizer's transformers_config in, it loads quantized; otherwise it's full bf16 moved to your device.

Why two names exist

Look at the class name and the display name and it starts to make sense: the node is literally TryOnOffModelNode - "load a model for either try-on or try-off" - and the pack just gave that one class two display names, plus a pair of run nodes to match. It's a mild naming mess, the kind of thing a small pack accrues while it's still being shaped. It costs you nothing: use whichever name you find in the workflow you loaded.

The practical guidance is the same as the other loader: pick cat-tryoff-flux for try-off, beta over alpha for try-on unless you're reproducing an old workflow, and if the graph you imported references "TryOff Model Loader" but you only see this one (or vice versa), just swap it in - it wires identically.

Install

Same pack, same steps:

cd ComfyUI/custom_nodes
git clone https://github.com/asutermo/ComfyUI-Flux-TryOff

or ComfyUI Manager → search "ComfyUI-Flux-TryOff", then restart. Needs diffusers>=0.32.2, transformers>=4.38.2, accelerate, numpy<2.0; add bitsandbytes if you quantize. First run downloads the model - a few GB, no HF token needed.

When to reach for it

Honestly, this is the node you'll use when a workflow someone shared drags in "TryOnOff Model Loader" and you'd otherwise have to hunt for the other name. It's a perfectly good loader - it just isn't a distinct one. Don't overthink the choice; the article you actually want on the loader details, the model options, and the speed realities is the one for TryOff Model Loader.