Simple Load FLUX

FLUX.1 is not one file. It's a diffusion model plus two text encoders (CLIP-L for the concise bits, T5-XXL for the long prose) plus the FLUX autoencoder VAE - and until you've loaded all of them correctly, you get nothing. Stock ComfyUI splits this across several loader nodes, which means hunting through text_encoders and vae folders, getting the exact filename right, and wiring three things by hand. SimpleSyrup.SimpleLoadFlux collapses it to one node: pick your diffusion model, everything else defaults to auto, and you get the standard model, clip, vae outputs.

The mechanism is the interesting part, because this is where "auto" stops being a convenience and becomes a correctness guarantee. FLUX.1's text encoders and VAE are revision-locked and checksum-pinned in the pack's catalog - CLIP-L and the FP16 T5-XXL from comfyanonymous/flux_text_encoders, and the FLUX VAE from the Comfy-Org repackaged files. When you leave a component on auto, the node finds the exact file if it's already on disk, and if it isn't, downloads it through HuggingFace with visible Comfy progress - and validates the download against the pinned SHA-256. A corrupt or wrong-version file is caught instead of silently poisoning every image you make. That's a genuinely better experience than "I downloaded a T5-XXL and my prompts look wrong."

Inputs:

• diffusion_model - the FLUX.1 model file. This node never downloads the diffusion model - you still install and select that yourself, from whatever source you got it (fp8 quantized on CivitAI or HuggingFace is the common route).
• diffusion_weight_dtype - precision for the diffusion weights, default default. FP8 uses less memory and can slightly change results; on a 12GB card that's often the difference between running and not.
• clip_l - auto, or pick an explicit CLIP-L file.
• t5_xxl - auto, or pick a T5-XXL. The tooltip warns the auto choice may download the large FP16 file - expect a multi-gigabyte download here, so budget for it and don't panic when a node suddenly starts pulling gigs.
• text_encoder_device - default puts both encoders where Comfy normally would; CPU saves GPU memory but makes prompt encoding slower. The tradeoff is real and it's the first lever to pull when the sampler is OOMing.
• vae - auto, or an explicit VAE file.

The clip output combines CLIP-L and T5-XXL into the dual-encoder CLIP object FLUX expects, so downstream it behaves like any other CLIP - feed it to a text encoder node and you're done.

Install: the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Artificial-Sweetener/SimpleSyrup
cd SimpleSyrup && pip install -r requirements.txt

or ComfyUI Manager → search SimpleSyrup → Install → restart. Keep ComfyUI current (v3 extension API). Requirements include huggingface-hub for the downloads. Troubleshooting in one line each: components don't appear in dropdowns → check the "Show downloadable models in loader dropdowns" setting, which controls whether known-not-yet-installed entries show up; download fails mid-way → checksum pin will catch it, delete the partial file and retry; prompts look subtly wrong → verify which T5-XXL got picked.