Recipe Model Loader

Recipes store model names, not model objects. That's the whole design: a recipe is data you can save, share, and edit - it can't carry a loaded model around. Recipe Model Loader (class RecipeModelLoader) is the node that bridges the gap: you feed it a recipe, tell it which model slot to load, and it resolves the model name into an actual MODEL, CLIP, and VAE you can wire into standard ComfyUI nodes.

How it works

Two required inputs drive everything:

• recipe_data - the recipe payload containing the model name strings.
• model - which slot to load: model_a, model_b, model_c, or model_d. Recipes can describe up to four model slots (image model, video model, whatever your pipeline needs), so this is how you say "give me the model from slot B."

There's also weight_dtype, a precision override for diffusion/UNet models - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. If you're on a VRAM-constrained card, fp8 is the lever that trades a little quality for a lot of headroom; it's ignored for checkpoints, which load as-is.

The three outputs are model, clip, and vae - standard types that plug into a KSampler, text encoders, and the decode side of your graph. The loader handles the resolver logic: it matches the recipe's model name against your installed checkpoints, diffusion models, and VAEs (including GGUF files if ComfyUI-GGUF is installed), so the recipe can travel between machines with different file layouts.

Why it exists

The recipe toolset has two ways to use recipes. The Recipe Renderer does everything internally - loads, samples, decodes - which is convenient but opaque. Recipe Model Loader is the manual alternative: it unpacks a recipe's model slot into ordinary sockets so you can build the graph your own way and only take the model-loading convenience from the recipe. It pairs naturally with Recipe Relay, which unpacks the rest of the recipe (prompts, sampler settings) into typed outputs - together they let a recipe drive a hand-built workflow rather than a fixed render template.

Installing

Part of the ComfyUI-Prompt-Manager pack:

cd ComfyUI/custom_nodes
git clone https://github.com/FranckyB/ComfyUI-Prompt-Manager.git
cd ComfyUI-Prompt-Manager
pip install -r requirements.txt

Or install via ComfyUI Manager (search "Prompt Manager"), then restart.

Gotchas

The failure mode that will actually bite you: the recipe names a model you don't have installed, and the loader can't resolve it. The node expects the recipe's model strings to match something in your checkpoints, diffusion_models, or vae folders - a recipe extracted on a machine with a different install won't just "work," you'll need to fix the name or install the model. And if the recipe's slot has no model name at all (an empty slot in a partial recipe), the load comes back empty, which downstream nodes will treat as an error. Check the slot you selected has a real model behind it.