⚡ Gigachad Model Loader

Gigachad Model Loader is the node for people who think a diffusion model isn't a single file - it's a set of memory trade-offs you get to choose. It loads a diffusion model (as opposed to a full checkpoint) and gives you the dials the stock loader hides: GGUF vs. safetensors, weight dtype, Sage Attention, and a Flux KV cache. It's the densest node in the pack, and the one whose feature names you should read before trusting.

The inputs, decoded

• model_name - everything in your diffusion-models folder.
• loader_type - diffusion_model (plain safetensors/pt) or GGUF. The GGUF path is the headline feature: pick a .gguf model here and the node quantizes loading through ComfyUI-GGUF.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2, fp16, bf16. fp8 is where the VRAM savings live; bf16 is the safe quality pick on RTX 40-series and newer.
• sage_attention - a dropdown with more options than the Checkpoint Loader's boolean: disabled, auto, then specific kernel selections (sageattn_qk_int8_pv_fp16_cuda, the _triton variant, _fp8_cuda, _fp8_cuda++, and two sageattn3 options). auto picks for you; the specific ones are for when you know your GPU wants a particular kernel. All require the sageattention package.
• flux_kv_cache - boolean. For Flux-family models, caches and reuses key/value tensors to cut memory and speed up generation at a small quality cost.

The single output is model (MODEL), ready for a sampler.

The GGUF gotcha

GGUF loading only works if ComfyUI-GGUF is installed. The node scans your custom_nodes folder for it at startup; if it's missing you'll see ComfyUI-GGUF not found - GGUF loading disabled in the console and the loader_type silently underperforms. That's a real trap if you follow a guide that says "just pick GGUF." Install the prerequisite:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF.git

then restart. Sage Attention has the same pattern - pip install sageattention or the feature logs a warning and skips. Both are optional accelerations bolted onto an otherwise dependency-free pack.

Install the pack itself

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/comfyui-gigachad.git

or install comfyui-gigachad via ComfyUI Manager and restart. No requirements.txt ships in the repo.

Should you use it?

If you already run ComfyUI-GGUF's own UnetLoaderGGUF and a separate sage patch, this node is a convenience wrapper around things you've already assembled - you won't gain anything but one node instead of three. If you're new to GGUF and fp8 and want a single place to experiment with memory settings, it's genuinely nice: you can A/B fp8 vs bf16 and sage on/off without rebuilding your graph. Just keep the two optional dependencies in mind, because a silent feature-disable is the one way this node will burn you.