Preface¶

When models in DSH need to generate images, the common practice is to call cloud APIs: there is quota consumption, content moderation, and assets do not leave the local machine. dsh-comfyui-studio takes a different path: the plugin calls your local or remote ComfyUI directly. Image generation and short video generation are processed by the local GPU. Parameters are modified in the control panel within the app, and after modification, they are written back to cordis.patch.yml. After restarting, the configuration persists.

It belongs to Lcxy0316, the repository is dsh-comfyui-studio, and the directory page categorizes it under “Model Inference”. The README is tagged with version 0.2.0 and the MIT license. Below is an introduction to how to install and use it.

I. What is it¶

One-sentence positioning: It installs a “ComfyUI Console” for DeepSeek Harness. Models (agents) generate images and short videos by calling local ComfyUI through three tools. The plugin does not preset any local paths; after installation, it is empty and needs to be filled manually.

It is worth noting that the plugin directory page you see is an independent site and has no official affiliation with DeepSeek / Huanfang; it is not an official app store.

II. The Three Provided Tools¶

After installation, models can directly call three types of interfaces.

comfyui_generate¶

Image generation interface. All parameters are optional; omitting them falls back to the defaults in settings. Therefore, a single call comfyui_generate({ prompt: "..." }) is a complete invocation.

comfyui_video¶

Video generation interface, uses Wan 2.1 text-to-video, returns two artifacts: an mp4 file and its first frame png. The reason is that the chat interface only renders images and does not play videos; the first frame is the one that can be seen directly.

comfyui_studio¶

Debugging and control interface, distinguished by action:

III. Installation and Enabling¶

Installation follows the order: “Get directory -> Put in profile -> Register -> Restart”.

1. Get the directory¶

The README provides a direct download zip method; after unpacking, you get the dsh-comfyui-studio directory:

https://github.com/Lcxy0316/dsh-comfyui-studio/archive/refs/heads/main.zip

2. Put into profile¶

cp -r dsh-comfyui-studio "$HOME/.dsh/profiles/desktop/node_modules/"

3. Register in package.json¶

Make two modifications in the profile’s package.json:

"dependencies": {
  "dsh-comfyui-studio": "file:./node_modules/dsh-comfyui-studio"
}

"dsh": {
  "profile": {
    "bundles": [
      "dsh-comfyui-studio"
    ]
  }
}

4. Restart DSH¶

The plugin’s built-in cordis.patch.yml will automatically take effect as a bundle layer; no user-layer additions are needed. Only when you need to override defaults, append to the end of the profile’s cordis.patch.yml:

- id: comfyui-studio
  name: dsh-comfyui-studio
  config:
    serverUrl: http://127.0.0.1:8188
    outputDir: D:/images

5. Four items to fill on first startup¶

The plugin does not preset any local paths; after installation, it is empty:

You can generate images without filling the first two items (as long as ComfyUI is already running), and the last two only affect one-click startup. The remaining fields (size, steps, prompts, etc.) vary by person, but you can run without changing them.

IV. Typical Usage: From Probe to Image Generation¶

First do a probe, then select the model, and finally generate images.

1. Open the settings page and first check the Connection group. serverUrl defaults to http://127.0.0.1:8188, requestTimeoutMs defaults to 600000 (10 minutes). Running 832×1216 / 28 steps on 4G VRAM takes about 90–120 seconds, so the default timeout is generous.
2. Click “Read Models” to pull a list from the server and select a checkpoint.
3. Fill in outputDir, which is an absolute path you have, for example D:/images.
4. Call comfyui_generate({ prompt: "..." }) to generate an image.

Common generation parameters: width, height, steps, cfg, sampler, scheduler, batchSize, seed. A common combo for SDXL Realism models is dpmpp_2m + karras + CFG 5–6 + 28 steps.

There are three layers of sources for prompts, with priority being “Tool call explicit > Prompt file > Settings default value”. The fields are positivePrompt, negativePrompt, positivePromptFile, negativePromptFile. The benefit of using prompt files is that changing words only requires editing the txt file, without touching the plugin or restarting.

In the Output group, outputDir empty means no saving to disk; saveToOutputDir defaults to true; filenamePrefix defaults to dsh; attachToConversation defaults to true, determining whether to embed the image back into the chat.

V. Hardware Presets¶

The settings page is arranged in 8 groups (Connection, Model, Generation Params, Prompt, Output, Video, Launcher, Advanced), plus four hardware presets:

One point to keep clear: applying a preset only changes the staged values on the page; clicking “Save Settings” writes them back.

Experience with 4G VRAM: Keep batch at 1, do not exceed resolution 1024×1408; if you need larger, generate first and then upscale (Hires fix uses 4x-UltraSharp, denoise around 0.35).

VI. Need an extra step for video¶

Wan 2.1 text-to-video is not a “single checkpoint”; it is split into three files, all of which need to be put into the ComfyUI model directory:

Direct connection to HuggingFace in China is generally not available; just change the domain to hf-mirror.com. The three files together are about 9.8 GB.

After placing them, click “Read Models” once, and then in the settings page “Video (Wan 2.1)” group, select the three items videoUnet, videoClip, and videoVae. Default values for the video group: videoWidth / videoHeight 480 / 832, videoFrames 49, videoFps 16, videoSteps / videoCfg 20 / 6.0, videoSampler / videoScheduler uni_pc / simple, videoShift 8.0.

VRAM reality: Wan2.1 1.3B can run on 4G, but it is slow; 480×832 / 49 frames / 20 steps takes minutes. First use frames=9, steps=4 to verify the pipeline, then increase from there. 1.3B is the only realistic choice for 4G VRAM.

VII. Troubleshooting¶

VIII. Who it’s for, what to note¶

Suitable for developers who already have local or remote ComfyUI and want models to generate images and videos locally without leaving the local machine, and are willing to fill in paths themselves in the settings page. People who don’t want to touch the local environment and just want to use ready-made cloud interfaces should not choose this plugin.

Two reminders. First, the plugin itself runs with the permissions of the current dsh process; you should check its source code and license before installing. Second, it does not pass through any cloud interfaces, so there is no quota consumption, no content moderation, and images and videos do not leave the local machine—this means the security boundary is entirely determined by your own ComfyUI environment.

IX. Links¶

• Plugin Directory Page: https://www.skillhub.cn/plugins/Lcxy0316/dsh-comfyui-studio
• GitHub: https://github.com/Lcxy0316/dsh-comfyui-studio