RedNode Rig Model

ComfyUI's problem with big one-panel node packs is always the same: the panel wants to load your model for you, and your model is weird. Maybe it's a GGUF. Maybe it's an INT8 W8A8 file. Maybe it's a merge you patched with three other nodes, or a video model, or something a pack wrapped in a union control. RedNode Rig Model is the pack admitting that and giving you the front door instead.

What it's for

The RedNode Studio Workspace runs your whole image on rigs - a rig being a model, its CLIP, its VAE, its sampler numbers, its LoRA set, and the kind of loader it uses. By default the Workspace loads those files itself, by filename, and it's genuinely good at it: it'll route a .gguf through ComfyUI-GGUF and an INT8 file through ComfyUI-INT8-Fast if you have them.

Rig Model is the escape hatch for everything else. Put it after your loaders, give it the rig's name, and the Workspace treats that rig as kind Your own nodes: your loader supplies the weights, and the Workspace carries on doing what it's actually good at - the Prompts tab, the camera stage, the LoRA stack, the Subject and Scene references, the Detailer passes, the grading, the save tree.

How it works

Three optional sockets in, three sockets out: model, clip, vae both ways. Nothing here is a pass-through with a bow on it, though. When the node runs, it registers those three objects with the Workspace under the rig name, and the Workspace applies that rig's LoRA set and camera LoRAs itself before your sampler ever sees the model. So the model coming out of the right of this node, during a rig run, is your loader's model plus the stack you configured on the LoRAs tab. Which is the whole trick: you get custom loading without giving up the LoRA management and the camera slider geometry that made you open the pack in the first place.

The registration is by name in a table, last writer wins. Two Rig Model nodes both called My rig will fight, and the pack prints a warning to the console saying which one it used. Only a node that actually ran in this queue counts, so a Rig Model you deleted or bypassed three sessions ago isn't silently resurrected.

One subtlety worth knowing before you spend an hour on it: the Workspace encodes this rig's prompts with the CLIP you hand it. On an ordinary SDXL or Flux setup that's unremarkable. On a Krea 2 rig it matters a lot - Krea 2's text encoder is Qwen3-VL with the vision weights, loaded through CLIPLoader with type krea2, and if you wire in some other encoder the identity and moodboard encoding that Krea 2 rigs depend on simply won't be there.

The console line worth watching is short: [RedNode Rig] 'My rig' ready: model, clip, vae. If the list is short, you left a socket empty.

Install

Nothing extra. The pack ships zero pip dependencies, which in an ecosystem where half the support threads are dependency conflicts is a small mercy.

cd ComfyUI/custom_nodes
git clone https://github.com/RedNodeAI/ComfyUI-RedNodeStudio.git ComfyUI-RedNodeStudio

Or find RedNode Studio in ComfyUI Manager, restart ComfyUI, and look under RedNode/Rigs. Optional packs like ComfyUI-GGUF, ComfyUI-INT8-Fast, Easy-Sam3 and SeedVR2 are only needed for other nodes - this one just needs your loaders.

Where people get burned

The name is the wire. All three rig nodes plus the Models tab entry have to agree. If they don't, the rig fails to load and the Workspace says "the RedNode Rig Model node for the rig 'X' did not run this queue", then falls back to a rig that does exist - which renders a picture with the wrong model, and that's how you lose twenty minutes.

Double LoRAs. Your rig's LoRAs are applied for you. Load the same LoRA again in your own chain and you get its strength twice, which is a soft, flattering kind of wrong that's easy to miss on faces.

Bypassing the head node. Rig Model has to run. Bypass it to "just test the sampler" and the chain has no starting point at all.

Forgetting the VAE. You can leave vae unwired and the node will still register - the console even tells you it registered model and clip only. The bill arrives later, in the Rig Result step, when a decoded picture has to be turned back into a latent and there's nothing to do it with.