Prompt file imported from Mooshieblob1/MooshieUI (.github/prompts/add-comfyui-node.prompt.md). Copyright stays with the author.

Add Custom ComfyUI Node (MooshieUI)

MooshieUI deploys its own ComfyUI nodes from src-tauri/src/comfyui/mooshie_nodes.py (embedded via include_str!, written to ComfyUI's custom_nodes/ at startup). Existing examples: MooshieSaveImage, MooshieFaceDetailer, MooshieSegmentDetailer.

Touchpoints (in order)

1. Python — src-tauri/src/comfyui/mooshie_nodes.py

class MooshieMyNode:
    """One-line summary of what the node does."""

    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "image": ("IMAGE",),
                "model": ("MODEL",),
                "vae": ("VAE",),
                "positive": ("CONDITIONING",),
                "negative": ("CONDITIONING",),
                "seed": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFFFFFFFFFFFF}),
                # sampler/scheduler dropdowns:
                "sampler_name": (comfy.samplers.KSampler.SAMPLERS,),
                "scheduler": (comfy.samplers.KSampler.SCHEDULERS,),
            }
        }

    RETURN_TYPES = ("IMAGE",)
    FUNCTION = "process"
    CATEGORY = "mooshie"

    def process(self, image, ...):
        # IMAGE tensors are [B, H, W, C] float 0–1
        return (result,)

Register at the bottom of the file:

NODE_CLASS_MAPPINGS = { ..., "MooshieMyNode": MooshieMyNode }
NODE_DISPLAY_NAME_MAPPINGS = { ..., "MooshieMyNode": "Mooshie My Node" }

2. Rust verification — src-tauri/src/comfyui/nodes.rs

Add the class name to REQUIRED_MOOSHIE_NODE_CLASSES so startup verifies ComfyUI actually loaded it (catches stale-server cases where files exist on disk but /object_info lacks the class until restart).

3. Workflow hookup — src-tauri/src/templates/

• New post-process chain → append_*_chain(result, params, image, seed) -> (String, u32) module, called from finish_workflow in mod.rs.
• finish_workflow chain order matters: upscale → facefix → segment → MooshieSaveImage. Insert new steps deliberately.
• Seed offsets: base seed for KSampler, seed+2 facefix, seed+3+i segments. Pick an unused offset.

Python node conventions

• Heavy imports (transformers, ultralytics) go inside methods, not module top — keeps node load cheap and the dependency optional.
• Model weights cache under folder_paths.models_dir subdirs (e.g. models/clipseg/); del model objects in finally to release VRAM.
• Sample with comfy.sample.sample(...) + comfy.sample.prepare_noise; preview via latent_preview.prepare_callback.
• Soft masks blend better than binary: return sigmoid/confidence values and let the composite weight per-pixel.
• Guard empty detections (mask is None, ys.numel() == 0) — return the input image unchanged, print a [MooshieMyNode] prefixed line for diagnosability (ComfyUI stdout, not Tauri logs).
• Python deps the node needs at runtime: ensure via installPipPackage("pkg==x.y.z") from the frontend before generation (see ensureFacefixPythonDependency in GenerateButton.svelte).

Verify

python -m py_compile src-tauri/src/comfyui/mooshie_nodes.py
cargo check --manifest-path src-tauri/Cargo.toml

- [ ] Class + NODE_CLASS_MAPPINGS + NODE_DISPLAY_NAME_MAPPINGS
- [ ] REQUIRED_MOOSHIE_NODE_CLASSES in nodes.rs
- [ ] finish_workflow / template chain wired (if workflow-facing)
- [ ] ComfyUI restarted when testing (deploy alone doesn't reload classes)