Recipe Relay

Recipe Relay (class RecipeRelay) is the bridge between the pack's recipe format and ordinary ComfyUI. A recipe is one RECIPE_DATA payload - prompts, LoRAs, sampler settings, resolution, model slot - and Relay unpacks it into individual typed outputs, plus lets you override anything on the way through. It's what you use when you want a recipe's convenience but your own graph's freedom: load a recipe, let it drive a workflow you built by hand instead of the fixed Recipe Renderer template.

How it works

It reads the recipe_data input and splits it into typed outputs: model, clip, positive, negative, vae, latent, image, mask, plus the sampler knobs - seed, steps, cfg, sampler_name, scheduler, denoise - and the geometry ones - width, height, batch_size, length - along with pos_prompt, neg_prompt, and lora_stack. Every one of those outputs also has a matching optional input, and here's the trick: any input you connect overrides the corresponding recipe value. Connect a seed INT and it replaces the recipe's seed; leave it unconnected and the recipe's value flows out. It's a pure passthrough bridge - nothing is modified or cached in the payload unless you use the dedicated override fields.

The model_slot selector (model_a–model_d) chooses which of the recipe's up to four model blocks you're reading and updating. The recipe-side writer inputs matter too:

• model_name - resolve a model name and write it into the recipe's model fields.
• family - override the slot's model family in the recipe data.
• model_data - take a structured payload from Recipe Model Picker and use it.

Models, CLIPs, and VAEs passed through are cached on the selected model block, so a hand-loaded model can be reused by the recipe without re-loading.

Why it exists

Two philosophies live in this pack. The Recipe Renderer says "give me a recipe, I'll handle everything." Recipe Relay says "give me a recipe, I'll give you all the parts, you assemble." Relay's audience is people who want recipes for reproducibility but refuse to give up their custom graphs - your own KSampler chain, your own upscaling, your own ControlNet. You extract a recipe, wire Relay to your existing graph, and the recipe fills in all the settings you used to set by hand.

Installing

Part of the 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 ComfyUI Manager → search "Prompt Manager", then restart.

Gotchas

The biggest confusion is the override behavior: unconnected inputs mean "use the recipe value," so if you want an override to apply, you must connect something to that input - a typed widget or a constant. And because nearly everything is optional, a recipe with empty slots will happily flow through and produce empty outputs; check that the recipe actually has the data you expect. Finally, don't confuse this with the Renderer: Relay does no sampling or decoding itself. If you wire it up and get no image out, that's working as designed - it's a plumbing node, not a renderer.