跪拜 Guibai
← All articles
Backend

WorkBuddy Taps Local ComfyUI for Credit-Free Text-to-Image Generation

By 当时只道是寻常_ ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Per-image credit fees disappear when image generation runs on a local ComfyUI instance through WorkBuddy's skill system, so the meter only tracks your GPU's power draw. The pattern also demonstrates how a skill-creator plugin can turn a plain-language workflow spec into a working, constrained skill without hand-written code.

Summary

WorkBuddy's built-in image generation costs at least 5 platform credits per image, a price that daily free allowances exhaust quickly. A workaround connects a local ComfyUI instance to WorkBuddy, letting the agent call a text-to-image workflow directly over its local API. The setup requires exporting the ComfyUI workflow as an API (not just saving it), installing the skill-creator plugin from ModelScope, and then dragging the exported JSON plus a detailed specification into WorkBuddy. The skill-creator automatically generates a ComfyUI text-to-image skill with parameter validation, strict image-size limits, error handling, and model unloading after each run.

Once installed, the skill triggers the local workflow through ComfyUI's HTTP endpoint at 127.0.0.1:8188. WorkBuddy can even optimize prompts before sending. The final step modifies an existing ink-wash sketch skill to call this new ComfyUI skill instead of the credit-based image model. The result is unlimited illustrations without spending platform credits, as long as the local GPU and ComfyUI remain running.

Takeaways
— Export a ComfyUI workflow as an API, not a saved workflow; only the API export exposes an external callable interface.
— Install the skill-creator plugin from ModelScope with `npx skills add https://www.modelscope.cn/skills/@anthropics/skill-creator`.
— ComfyUI must be running when WorkBuddy generates the skill; otherwise the creation fails.
— The generated skill restricts image dimensions to a fixed set of aspect ratios, from 1:1 at 1328x1328 to 16:9 at 1664x928.
— WorkBuddy calls the local ComfyUI workflow over HTTP at 127.0.0.1:8188, producing images without consuming platform credits.
— Modifying an existing skill to call the ComfyUI skill replaces the credit-based drawing step with free local generation.
Conclusions

The credit meter is the real driver; the free daily allowance is trivial for image generation, so a local fallback offers a practical escape hatch for heavy users.

The skill-creator plugin converts a natural-language specification into a working skill, but the spec must be detailed enough to enforce parameter locking, size validation, and error handling for the result to be reliable.

This workflow depends on a local GPU and a continuously running ComfyUI, so it is a self-hosted alternative to managed API pricing, not a universal solution for every user.

Concepts & terms
ComfyUI
A node-based tool for building and running Stable Diffusion image-generation workflows locally. It provides an HTTP API that external programs can call to execute a saved workflow.
WorkBuddy
An AI agent platform (similar to Claude or ChatGPT) where users define reusable 'skills' that the agent can invoke during a conversation. Skills can call external tools or APIs.
Skill-creator
A plugin skill that automatically builds, modifies, and optimizes other skills based on a natural-language specification, following a defined set of steps and standards.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗