DEV Community

AI-KSK
AI-KSK

Posted on

Writing Your Own ComfyUI Node With VS Code, Cline and Doubao Seed 2.0 Code

#comfyui #ai #tutorial

Every ComfyUI user eventually hits the same wall: the node you need does not exist, or the one that exists quietly recompresses your image on the way out. This episode's answer is to stop searching and write the node. The author does it as a self-described beginner, with VS Code, the Cline extension and Doubao Seed 2.0 Code, and the example is a captioning helper that saves an image and a matching text file.

Evidence note: this episode ships no workflow JSON and no node package. What follows is a transcription of the author's on-camera walkthrough, including the bug he hits and how he fixes it.

When this is useful

  • A step in your pipeline needs a node nobody has published, and the nearby options are almost right.
  • You are chaining nodes for batch captioning and one of them degrades image quality.
  • You want to try AI-assisted coding without paying for it: the whole stack here runs on free quota.

What is different about this approach

  1. The node is generated, not hand-typed. The author writes the requirement in plain language and lets Seed 2.0 produce the implementation, then iterates on the result.
  2. The editing happens where ComfyUI can see it. The project folder is created directly under ComfyUI's custom_nodes directory, so a finished node can be tested in the same graph immediately, without copying files around.
  3. The example is a real gap in the author's own workflow. Existing captioning helpers load the image, then output a compressed copy. The custom node keeps the image and writes a same-named text file beside it.
  4. The free tier is the point. Volcano Ark gives new accounts a per-model token allowance, and identity verification raises the ceiling to the advertised maximum of 5 million tokens.

Step by step

1. Install VS Code. Download from the link in the video description, install, open.

2. Get an API key. Register on Volcano Ark through the description link (the author notes monthly coupons), complete identity verification, open the model marketplace, pick Doubao Seed 2.0 Code, open API access and create an API key.

3. Install Cline. In VS Code, open the extensions panel, search Cline, install, then open its settings.

4. Point Cline at Seed 2.0. Choose the OpenAI-compatible provider option, paste the Base URL, paste the API key you just created, and paste the Model ID. Confirm, and Cline is now calling Seed 2.0 Code.

5. Create the project folder under ComfyUI's custom_nodes, so testing does not require a second copy.

6. Describe the node. The author's requirement: a node that assists batch captioning and saves both the image and a text document. Seed 2.0 returns an implementation; accept the diff and let it write the files.

7. Restart ComfyUI and look for the node. It appears as Save Image with Text, and its own docstring explains what it does, which is a useful side effect of generating the code with the requirement attached.

8. Wire it into the batch captioning graph and test. Insert it where the old captioning helper sat, run one batch, and check the outputs.

9. Fix what breaks. In the recording the first run is wrong: the generated text file does not share the image's filename. The author restates the requirement as "same name, different suffix", and the second iteration is correct.

10. Keep it. The node now lives in custom_nodes and survives restarts, which is the whole point of owning it rather than borrowing it.

Parameter basis

No workflow JSON exists for this episode, so there is nothing to read parameter values from. The only quantitative claims in the material are the free-quota figures: a per-model allowance on registration and an advertised maximum of 5 million tokens after verification. Those are platform terms quoted by the author, not settings of the node.

Limits

  • No JSON, no node package. The episode documents a process; the resulting node is not distributed in the folder.
  • The demo is one node. A captioning helper is a small, well-bounded task; the approach is not evidence that Seed 2.0 can generate a complex multi-input node unsupervised.
  • The first generation was wrong. It took a second prompt to make the text file match the image name, which is the realistic failure rate to expect.
  • Free quota is finite and platform-controlled. A per-model allowance and a verified-account ceiling are terms the platform can change; nothing here is a permanent free tier.
  • The author is a beginner at this by his own account, so the walkthrough reflects beginner friction rather than best practice for node architecture.

Episode resources

Top comments (0)