What problem does it solve?
A ComfyUI render that completes but looks wrong (artifacts, wrong pose, ignored ControlNet, degraded upscale) is hard to fix by re-running the whole graph blindly. This Skill localizes the first bad stage by rendering only up to a chosen output node and inspecting what each stage actually produces.
Core Features & Use Cases
- Partial execution probing: Uses
panel_run with to_node_id so only the target output node and its upstream branch render, making each probe fast and cheap.
- Preview taps for non-output points: Adds temporary PreviewImage, VAEDecode, or MaskToImage taps to inspect latents, masks, preprocessor maps, and intermediate images.
- Bisection workflow with symptom guide: Maps symptoms (ControlNet ignored, inpaint bleeding, refiner degradation, bad quality) to the right first probe point, then bisects until the faulty node is found.
- Use Case: A txt2img plus upscale pipeline produces a blurry final image. Preview the base latent before the refiner, confirm it is good, then fix the refiner stage's sampler settings and confirm with another run-to-node before the full render.
Quick Start
Ask the agent to debug why the final image looks wrong by running the graph only up to a preview tap on the suspected stage and inspecting the delivered output.