# ComfyUI-zhenzhen: ComfyUI nodes pointed at a resale API

> ComfyUI-zhenzhen is a set of ComfyUI custom nodes that call a third party model API rather than running models locally, covering text to image, image to video, music and upscaling across a long list of models. It is a well maintained wrapper with a detailed update log, but it depends entirely on one reseller, and its own documentation says the service is not for commercial use.

**T8mars/Comfyui-zhenzhen** — https://ai.t8star.org/register?aff=dP7j Comfyui的各种平价API调用节点，支持seedance2.0（满血版，支持真人，不排队）

- Repository: https://github.com/T8mars/Comfyui-zhenzhen
- Stars: 754 · Forks: 50
- Language: Python
- License: Apache-2.0
- Published: 2026-10-21 · Updated: 2026-10-21 · Language: en
- Canonical page: https://hysenlabs.com/projects/t8mars-comfyui-zhenzhen

## Nodes that call an API instead of running a model

ComfyUI-zhenzhen is a ComfyUI custom node pack whose nodes call a hosted model API. The description in the repository says it plainly: various affordable API calling nodes for ComfyUI, with support for seedance 2.0 in a full version that supports real people and does not queue.

The project is written in Python, published to the Comfy Registry under the publisher id t8star with the display name Comfyui-zhenzhen, and licensed Apache-2.0. The package metadata calls it zhenzhen, version 2.4.8, and the project documentation lives in a wiki on GitHub.

What this means in practice is worth being precise about, because it is different from most ComfyUI node packs. A typical node pack wraps a local model, running inference on your own GPU. This one wraps somebody else's inference. The graph looks the same, you drop a node onto the canvas, give it a prompt, connect an output, but the compute happens on the service's machines and the result comes back over HTTP.

That has an upside and a downside. The upside is access to models that would be impractical to run locally, including video models, at the cost of a per-call price. The downside is that the node is only as good as the API behind it, and the API belongs to one company.

The repository layout reflects a large, actively maintained node surface rather than a single integration. There are per-family modules such as seedance_low_price_nodes.py, gpt_image_25_nodes.py, topaz_low_price_nodes.py, midjourney_low_price_nodes.py, suno_workshop.py, fal_batch_nodes.py, qwen21_animate_low_price_nodes.py and nb21_flux_vidu_low_price_nodes.py, alongside shared plumbing in zhenzhen_http.py, media_download.py, config_store.py, utils.py and execution_seed.py, and workflow JSON files for individual models.

## Four entry points, and the difference between them

The README opens with a navigation table, and the four rows matter because they are not the same service.

The first two are the operator's own. The domestic edition, described as the affordable AI shop for domestic users and domestic models first, mainly calls domestic models, is a non-profit site that keeps only 5 to 10 percent for site maintenance, and prices domestic models at roughly 75 to 80 percent of the overseas prices. The overseas edition, the AI workshop, mainly calls overseas models and also carries domestic ones, costs more overall, and has no price advantage on domestic models.

The other two are RunningHub API keys, offered in domestic and overseas variants. The domestic key is aimed at users who need to adapt more AI applications and can try the newest models; the overseas key is described as supporting overseas models with looser review.

So there is a real choice to make, and it is not only about price. Which models you can reach, how strictly the content is reviewed, and where the requests are served all differ between the options.

The node pack supports both of the operator's channels, and the update notes distinguish them, with entries naming which channel a new node targets. The pricing philosophy is stated plainly as well: the site was set up to provide affordable API access to fans, is not run for profit, has a margin under 10 percent, and after tax and development and server costs has almost no profit.

## How a node actually works: upload, submit, poll, download

The update log describes the mechanics in enough detail to understand what a node does, and they are consistent across families.

A typical video node takes local material or a public URL, uploads it, submits a task, polls the task, and downloads the result when it completes. The log entries mention a 10-way concurrent video submission framework that new nodes plug into automatically, along with standard fixed and random seed caching controls, an API key retrieval button for the domestic channel, and dynamic showing and hiding of inputs depending on the selected model.

That model-switching behaviour is worth noting. When a node supports several models, only the controls that apply to the chosen model are shown, and the log for one entry notes that the overseas models do not submit enable_thinking while the frontend only shows that control for two standard Global models. That is the kind of detail that only appears when a node has been tested against the live API rather than assembled from documentation.

Download handling is where the engineering shows most clearly. One entry records that two Global tasks failed at the TLS handshake stage because of an old Hong Kong COS domain, and that the fix was a fallback to the official new Tencent COS domain, matching only the myqcloud.com hosts, preserving the signed path and query parameters and keeping certificate verification on. Another entry records that after a connection or TLS interruption during video streaming download, the route switches automatically to a path with no environment proxy, and that the final error message shows only the error type rather than the full result address.

Both entries are about protecting the download of a paid result, which tells you where the fragility of an API-based workflow lives.

## The update log doubles as a test record

The most distinctive thing about this repository is the update log, and it is worth reading as a testing record rather than as a changelog.

Each dated entry ends with regression numbers. One reports a complete offline regression of 220 out of 220 passing, with a plugin probe confirming 165 ordinary nodes, 130 concurrent submission nodes, 297 merged nodes and 32 concurrent validation agents. An earlier entry reports 219 out of 219, and another 215 out of 215, with the node and probe counts rising in step.

Alongside that there are the real API checks. The 2026-08-30 entry records that four Prime models completed generation and download as decodable MP4 files, with the resolutions, frame rates and frame counts listed for each, and that the flash node covered text to video, first frame, three image references and reference video with upload, submission, polling, download and MP4 decoding all performed. The 2026-08-24 entry records that the domestic models downloaded valid MP4s while the two Global tasks hit the TLS failure that was fixed the following day.

The counter style is unusual and mostly a good sign. It tells you the project distinguishes between tests that run offline against a stubbed API and checks that go against the live service, and it records the failures as precisely as the successes, including a case where the upstream returned a failure at the end of the task and the node preserved the correct protocol and passed the error through clearly.

## Dependencies, registry metadata and the working files

The packaging is straightforward and the dependency list tells you what the nodes do at the edges.

The pyproject file declares the project as zhenzhen, version 2.4.8, with the Apache licence file, and the dependencies are:

```python
dependencies = ["aiohttp", "aiohttp-cors", "GitPython", "numpy", "Pillow", "requests", "torch", "GitPython", "matrix-client", "transformers", "huggingface-hub", "psutil", "nest_asyncio", "openai"]
```

The requirements file adds opencv-python on top of that list, which is what the image and video handling nodes use for frame and media work. GitPython appears twice in the dependency list, and the openai client is present for the endpoints that speak that protocol.

The Comfy Registry metadata is in the same file, with a PublisherId of t8star, the display name and the icon, and a commented-out requires-comfyui line for a version constraint, so no minimum ComfyUI version is currently declared. The URLs point at the repository, the wiki for documentation and the issue tracker for bugs.

The top level of the repository is the clearest map of the project. There are per-family node modules as well as Comfly.py and ComflyConcurrent.py for the submission framework, api_settings_compat.py, config_store.py for settings, execution_seed.py for the seed controls, media_download.py for the download path, suno_workshop.py, a tests directory, a workflow directory with per-model JSON files, a web directory, a docs directory, a certs directory and pic for the images used in the README.

## The commercial framing, and the tax question

The README is direct about the commercial side, and an adopter has to read it rather than assume.

The site was opened with the aim of giving fans affordable API access to the newest overseas models, not for profit. The margin is under 10 percent, and after normal tax, development, maintenance and server costs there is almost no profit, so the site is not run commercially, has no room to negotiate prices, and does not support commercial secondary development. It serves fans only.

There is a second, less obvious condition. The README says the number of invoices per month is limited, and that users have to bear all the taxes themselves at 5 to 25 percent. For anyone using this to produce paid work, that is a real cost line and a compliance question, and it is the operator's own statement of the arrangement rather than an inference.

The register links are referral links with an aff or inviteCode parameter, and the entry table leads to a sign-up page rather than to a documentation site. The tutorials are hosted on the operator's own pages and on YouTube, including a Chinese-language playlist for the full model walkthrough, and there is a Windows integration package distributed through a file sharing link.

The practical reading is that this is a community service with a commercial backend, wrapped in a genuinely well maintained client. The code quality and the business model are separate questions, and an organisation evaluating it has to answer both.

## Conclusion

Use ComfyUI-zhenzhen if you want to drive current image and video models from ComfyUI without local hardware, and you are comfortable routing your prompts and uploads through a reseller service. Do not adopt it for commercial production, because the project states the service is not for commercial use and offers no room for negotiation on price, and do not adopt it if you need data residency, since every request goes to a third party. Before you install, read the pricing and tax notes, decide which of the four entry points suits you, and expect to work from Chinese documentation, with the latest release v4.9K from 2026-10-12 and the last push on 2026-10-13.

## FAQ

### What is ComfyUI-zhenzhen?

It is a set of ComfyUI custom nodes that call a third party model API for image, video, music and upscaling tasks, rather than running models locally. The project is published to the Comfy Registry under the publisher id t8star and is licensed Apache-2.0.

### Which models and channels do the nodes support?

The repository carries separate node modules for seedance, GPT image 2.5, Flux 3, Topaz, Midjourney, Suno, fal batch calls, Qwen animate, Doubao and Vidu workflows, and the update log names further additions. The operator's own service has a domestic edition that mainly calls domestic models at lower prices and an overseas edition that mainly calls overseas models, plus RunningHub API key options in both versions.

### Is the service suitable for commercial work?

The README states that the site is not run for profit, has no room to negotiate prices, does not support commercial secondary development, and that users must bear taxes of 5 to 25 percent themselves with a limited number of invoices per month.

### How does the project test its nodes?

The dated update entries report an offline regression count that rises across releases, along with a plugin probe that counts ordinary nodes, concurrent submission nodes and merged nodes, and they also record live API checks with the resulting file resolutions, frame rates and frame counts, including the cases that failed.

## Sources

- [Issues](https://github.com/T8mars/Comfyui-zhenzhen/issues)
- [License: Apache-2.0](https://github.com/T8mars/Comfyui-zhenzhen/blob/main/LICENSE)
- [README](https://github.com/T8mars/Comfyui-zhenzhen/blob/main/README.md)
- [Releases](https://github.com/T8mars/Comfyui-zhenzhen/releases)
- [T8mars/Comfyui-zhenzhen on GitHub](https://github.com/T8mars/Comfyui-zhenzhen)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/t8mars-comfyui-zhenzhen