Rebel Load Diffusion 🌌

Modern split-model workflows - Flux, SD3, Wan, and friends - don't load a monolithic checkpoint. They load the diffusion transformer separately (into the diffusion_models folder), the text encoder separately, and the VAE separately. Rebel Load Diffusion 🌌 is the animated version of that first step: it wraps ComfyUI's UNETLoader machinery but points its file dropdown at the diffusion_models folder, so it's the node you reach for when your model file lives there instead of in checkpoints. Same loader under the hood, matrix rain on top.

How it works

Under the animated body it's a thin pass-through to core ComfyUI's UNETLoader - the same code path the stock "Load Diffusion Model" node uses. The pack's wrapper grabs the native loader's input definitions, pins unet_name to the diffusion_models directory listing, and delegates the actual model load straight through. That means every weight-dtype option the core loader supports is available here, and when ComfyUI improves the loader, this node inherits it. The animation is the usual frontend overlay: pure decoration, zero effect on loading, zero VRAM cost.

The inputs that matter

• unet_name - the diffusion model file, from whatever's in models/diffusion_models. This is the dropdown that makes the node what it is; it will never show your checkpoints folder.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. This is the lever that actually matters on low-VRAM cards: fp8 roughly halves the model's memory footprint with near-zero visible quality loss, which is why it's the default recommendation for big models like Flux. On a card where the fp16 file barely fits, fp8 is often the difference between running and OOMing.
• theme / color_scheme - the pack's 12 animations and 14 palettes. Decoration.

The output is a single MODEL, wired into your KSampler (and into a LoRA loader if you're stacking LoRAs on top).

Installing it

One clone, no dependencies, no model files from the pack itself:

cd ComfyUI/custom_nodes
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 actually draws. Find it under Rebel AI → Animated Core. ComfyUI Manager can install the pack if you search "Rebels_Animated_Nodes".

Where people get burned

• "My file isn't in the dropdown." It's looking at diffusion_models specifically. Drop the model there (or use the Rebel UNET nodes that target the unet folder) and restart ComfyUI so the folder gets re-scanned.
• OOM at default, fine at fp8? That's the expected shape of things on an 8–12 GB card with a big transformer. Use fp8_e4m3fn and consider a quantized file before you blame the node - the wrapper is doing the exact same thing the core loader does.
• No animation on Nodes 2.0. The pack draws on the legacy LiteGraph canvas; with ComfyUI's newer frontend, expect the model to load fine while the backdrop goes missing.
• Red node / missing classes on load. The pack isn't installed or wasn't restarted after the clone. Confirm it's in custom_nodes and restart.

If your whole pipeline is modern split models, this is the loader to make your workspace feel cohesive - it's the exact tool, reskinned, for the exact folder your Flux-class models live in.