Rebel UNET Loader (GGUF) 🧠

If your card has ever said out of memory next to a 12B-parameter transformer, you know why GGUF exists: quantized weights that trade a little precision for a lot of VRAM headroom, letting big models run on cards that could never hold them at fp16. Rebel UNET Loader (GGUF) 🧠 is the pack's animated wrapper for ComfyUI's UnetLoaderGGUF node - the one that reads .gguf files from your unet folder. It's the low-VRAM path, dressed up to match the rest of your Rebel workspace.

How it works

This wrapper is slightly smarter than the pack's other loaders, because GGUF support is a third-party add-on rather than core ComfyUI. The node checks at load time whether the UnetLoaderGGUF class is registered (which requires the separate ComfyUI-GGUF pack from city96); if it is, the node uses it and inherits its real inputs; if not, it falls back to the plain UNETLoader so your server still boots instead of crashing. That defensive fallback is the pack's "crash-proof" loading in action - but it also means the honest read: the GGUF magic lives in ComfyUI-GGUF, not here. This node is the pretty front door to someone else's engine.

The inputs that matter

• unet_name - the .gguf file, listed from the unet folder.
• weight_dtype - default, fp8_e4m3fn, or fp8_e5m2. With GGUF files, the quantization is baked into the file itself (Q4, Q5, Q8 and so on), so this matters less than it does on a full-precision loader - pick default and let the file's own quant do the work.
• theme / color_scheme - the pack's usual 12 animations and 14 palettes.

The output is a single MODEL socket, feeding your sampler like any other model input. And a quick pointer for picking files: Q8 is basically fp16 at half the size, Q5 is the last tier before visible loss, Q4_K_M is the accepted 12 GB compromise - if the model fits at Q8, just use it.

Installing it

Two installs, and this is the gotcha. First the GGUF engine, then the wrapper:

# 1) the actual GGUF node pack (city96/ComfyUI-GGUF)
cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF.git
cd ComfyUI/custom_nodes

# 2) the Rebel wrapper
git clone https://github.com/RealRebelAI/Rebels_Animated_Nodes.git

Restart ComfyUI, then hard refresh the browser (Ctrl + F5 / Cmd + Shift + R) so the cached frontend JS reloads and the animation draws. The pack itself has no requirements.txt and downloads no models - your .gguf files are the only downloads.

Where people get burned

• "The Rebel GGUF node is red!" That's the pack's own documented troubleshooting entry: you don't have ComfyUI-GGUF installed. Either install it, or (per the README's own advice) delete the red node and use Rebel Load Diffusion with a regular model instead.
• The node exists but loads nothing GGUF-specific. Without ComfyUI-GGUF present, this node silently falls back to the plain UNET loader - your server won't crash, but you also won't get quantized loading. If you installed only the Rebel pack and expected GGUF support, that's the mismatch.
• Animations flat under Nodes 2.0. The pack renders on the legacy LiteGraph canvas; the new frontend can drop the backdrop while the load still works.

The takeaway: this is the right node if you're already running GGUF quantized models and want a consistent look across your graph. It is not the thing that makes GGUF work - install ComfyUI-GGUF first, or the fancy wrapper is wrapping an empty room.