ClownModelLoader

ClownModelLoader is a one-node loader: it pulls in your diffusion model, up to four CLIP/text encoders, and a VAE, and hands out all three (model, clip, vae) from a single box. On a modern multi-encoder setup - Flux with its dual encoders, SD3.5 with CLIP-L + CLIP-G + T5 - that's a lot of separate loader nodes collapsed into one, plus fp8 weight casting baked in for when VRAM is tight.

The reason a sampler pack ships its own loader is convenience and consistency. RES4LYF workflows already have a dozen specialized nodes on the canvas; folding the whole load stage into one keeps the graph readable, and having the dtype and encoder wiring in one place means fewer mismatches. Nothing here is unique magic - it's the stock loaders wrapped together - but the ergonomics are real, and its click-through rate suggests people go looking for exactly this.

How it works

You point it at a model file, choose a weight dtype, select which text encoder(s) and type to load, and pick a VAE source. It runs the equivalent of the separate UNET/checkpoint, CLIP, and VAE loaders and returns the three standard outputs. The CLIP slots can also reuse the checkpoint's built-in CLIP rather than a separate file.

The inputs and outputs that matter

• model_name - the model/checkpoint to load (the dropdown is populated from your models folder).
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. The fp8 options cut VRAM roughly in half for big models like Flux at some quality/precision cost; fp8_e4m3fn_fast is the speed-leaning one on supported GPUs.
• clip_name1_opt…clip_name4_opt plus clip_type - which encoders to load and how to interpret them. clip_type covers the whole modern zoo (flux, sd3, ltxv, hunyuan_video, cosmos and more); .use_ckpt_clip reuses the checkpoint's own CLIP.
• vae_name - a VAE file, .use_ckpt_vae to take the one baked into the checkpoint, or a TAESD preview VAE.

Outputs are model (MODEL), clip (CLIP) and vae (VAE) - the three you wire into the rest of a normal graph.

How to install it

Comes with RES4LYF. ComfyUI Manager: search RES4LYF, install, restart. Manual:

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

Portable builds use the embedded pip. Restart, hard-refresh (F5).

Common issues

The commonest headache is the CLIP configuration. If you get gibberish or a "text encoder mismatch" error, the encoder type or the number of encoders is wrong for the model - Flux wants CLIP-L + T5, SD3.5 wants three, and picking .use_ckpt_clip on a model whose file doesn't contain CLIP won't work. Match the encoders to the architecture. Second, fp8 is a real trade: it saves memory but a few models are visibly softer in fp8, and fp8_e4m3fn_fast only helps on GPUs with the hardware path for it - if you see no speedup, that's why. And the usual loader gotcha applies: if model_name is empty, ComfyUI can't see your files, which means they're in the wrong folder or you need to refresh the node list.