FluxLoader

Setting up a Flux workflow normally means a small pile of loader nodes: a UNet/diffusion-model loader, a dual CLIP loader for the two text encoders, a VAE loader, and - if you're doing Redux/style-transfer work - a CLIP Vision loader and a style model loader on top. FluxLoader collapses that pile into one node. Pick your model, and it hands you the model, CLIP, VAE, CLIP Vision, and style model out of a single box.

It exists because RES4LYF is heavily invested in Flux, and in particular in the style-transfer and faceswap features the author built on Flux Redux. Those pipelines need the vision and style-model outputs that a plain checkpoint loader doesn't give you, and wiring five separate loaders every time gets old. This is the convenience loader that keeps a Flux graph tidy.

How it works

You choose a model and a weight precision, and the node loads the components and exposes them as typed outputs. The CLIP, VAE, and other slots can either come bundled from the checkpoint or be pointed at separate files, so one node covers both the all-in-one checkpoint case and the mix-and-match case.

The inputs and outputs that matter

The two you'll actually touch:

• model_name - your Flux model, picked from your local models folder. (The dropdown is populated from what you have installed, so it reads empty until you've got Flux files in place.)
• weight_dtype (default / fp8_e4m3fn / fp8_e4m3fn_fast / fp8_e5m2) - the precision the model loads at. default is full precision and the heaviest on VRAM; the fp8 options roughly halve the model's memory footprint at a small quality cost, and fp8_e4m3fn_fast trades a bit more accuracy for speed. If Flux won't fit, this is the knob.

The rest - clip_name1, clip_name2_opt, vae_name, clip_vision_name, style_model_name - default to using the checkpoint's bundled components (.use_ckpt_clip, .use_ckpt_vae) or to none, and you only override them when you're supplying separate CLIP/VAE/vision/style files.

Outputs: model, clip, vae, clip_vision, and style_model - wire the first three into your sampler and encoders as usual, and the last two into a Redux/style-guide setup when you're doing style transfer.

Installing it

ComfyUI Manager: search RES4LYF, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF/
cd RES4LYF
pip install -r requirements.txt

pip install in your venv or portable Python, restart, hard-refresh (F5). The node itself downloads nothing - but you still need the actual Flux model, CLIP encoders, and VAE files in your models folders for the dropdowns to populate.

Where people get tripped up

The most common confusion is an empty model_name dropdown, which just means you haven't got the Flux files where ComfyUI looks for them yet - this loader doesn't fetch anything, it only lists what's already installed. Get Flux, its text encoders, and its VAE into the right folders first.

Second, the CLIP Vision and style-model outputs are for Redux-style workflows specifically. If you're doing plain text-to-image, leave those unwired and set to none - they're extra machinery you only need when you're feeding a reference image through Flux Redux for style or faceswap. And if you're on a tight VRAM budget, reach for an fp8 weight_dtype before you start disabling other things.