Use ModelSite in ComfyUI

ComfyUI is a node-based tool for image and video generation. ModelSite models can be used in it in two ways.

If you work in the ComfyUI interface, use Plan B. Use Plan A only when you drive ComfyUI from code and want to reuse ComfyUI's built-in nodes and the template workflows that ship with it.

Both plans bill at ModelSite prices, and every request shows up in your usage log like any other call. Keys look like sk-ms-...; create one in the console and never paste a real key into a workflow file, a script you share, or a screenshot.

Plan A: ComfyUI's own API nodes, pointed at ModelSite

1. Start ComfyUI with the ModelSite address

python main.py --comfy-api-base https://api.modelsite.ai

--comfy-api-base tells ComfyUI's built-in API nodes where to send their requests. Tested with ComfyUI 0.42.0.

2. Submit a workflow and pass the key

Build or open the workflow in ComfyUI, then use Workflow → Export (API) to save it in API format. The template workflows under Templates → API (for example api_seedance2_0_mini_t2v, api_wan2_7_t2v, api_happyhorse1_1_t2v) work as a starting point.

Send it to ComfyUI's /prompt endpoint with your ModelSite key in extra_data.api_key_comfy_org. ComfyUI forwards that value to ModelSite as the X-API-KEY header.

import json, os, time, requests

COMFY = "http://127.0.0.1:8188"          # your local ComfyUI
KEY = os.environ["MODELSITE_API_KEY"]    # sk-ms-...

with open("workflow_api.json") as f:     # exported with "Export (API)"
    prompt = json.load(f)

r = requests.post(f"{COMFY}/prompt", json={
    "prompt": prompt,
    "extra_data": {"api_key_comfy_org": KEY},
})
r.raise_for_status()
prompt_id = r.json()["prompt_id"]

# Video models take minutes: poll until the job shows up in history.
while True:
    h = requests.get(f"{COMFY}/history/{prompt_id}").json()
    if prompt_id in h:
        print(json.dumps(h[prompt_id]["outputs"], indent=2))  # files are in ComfyUI's output folder
        break
    time.sleep(5)

3. What works

Checked by running the template workflows end to end against a real ModelSite account, with the cheapest settings:

ComfyUI's nodes send dated model names such as dreamina-seedance-2-0-260128 or seedream-4-5-251128. ModelSite maps these to the matching model, so you do not need to change them.

4. What does not work with Plan A

• Nodes that upload files first. Image to video, reference-to-video, first-and-last-frame, video edit and similar nodes ask ComfyUI to upload your image or video before generating. ModelSite does not offer that upload step, so these nodes fail. Use Plan B for image to video.
• Model families ModelSite does not sell, such as Kling, Vidu and Luma. Their nodes have nothing to connect to.
• Models in ComfyUI's lists that ModelSite does not have: Wan 2.5 text to image, Seedream 5.0 Pro, Qwen Image 3.0 Pro, Seedance 2.0 Fast and Seedance 2.5. The request is refused.
• New models appear in ComfyUI's lists only when ComfyUI ships a new release. Plan B does not have this limit.

Plan B: the ModelSite node pack

The pack adds nodes that call ModelSite directly. There is no ComfyUI account, no sign-in gate and no launch argument.

1. Install

The pack is listed in the ComfyUI Registry as Modelsite. In ComfyUI Manager, open Custom Nodes Manager, search for Modelsite and click Install, then restart ComfyUI.

With the comfy CLI:

comfy node install comfyui-modelsite

Or clone the repository into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/binoctal/comfyui-modelsite

Restart ComfyUI. The pack needs no extra Python packages.

2. Set your API key

Open Settings → Modelsite → Connection and paste your key into API key. The key is stored on your machine only. It is never written into workflows or into generated files, so you can share a workflow without sharing the key.

Or set the MODELSITE_API_KEY environment variable before you start ComfyUI. The API address defaults to https://api.modelsite.ai; change it only if you were given a different one.

3. Add nodes

Double-click the canvas and search for Modelsite. The nodes are under Modelsite/Video and Modelsite/Image:

Connect the output to ComfyUI's standard Save Video or Save Image node. Video nodes also output task_id and usage (the usage numbers the API returned).

4. Pick a model

The model dropdown is filled from ModelSite's live model list when the page loads, so a newly added model shows up after you reload the page. If you need one right away, type its ID into custom_model; it overrides the dropdown.

Reference videos and audio are not supported yet.

Checked end to end in the ComfyUI interface with a real account: text to video on Seedance 2.0 mini, Wan 2.7, HappyHorse 1.1 and MiniMax H3; image to video on Seedance 2.0 mini, Wan 2.7, HappyHorse 1.1 and MiniMax H3; Seedream 5.0 lite text to image and Seedream 4.5 image to image. Other models in the table use the same nodes but were not each run.

5. Parameters worth knowing

• resolution: Wan and HappyHorse accept 720p and 1080p. Seedance accepts 480p, 720p and 1080p. MiniMax H3 accepts 768p (H3 Max also 480p). An unsupported value is refused before anything is sent.
• duration: Seedance 4–15 s, Wan 2–15 s, HappyHorse 3–15 s, MiniMax H3 4–15 s (H3 Max 5–15 s).
• seed: -1 lets the model choose. The default setting makes every run different, so ComfyUI never reuses an old result for a paid call.
• generate_audio works for Seedance only. negative_prompt works for Wan and HappyHorse only. Other models ignore them.
• ratio applies to text to video. In image to video the result follows your first frame.
• HappyHorse image to video takes a first frame only; connecting a last frame is an error.

6. Example workflows

Drag a file onto the ComfyUI canvas. Image workflows read example.png from ComfyUI's input folder; replace it with your own image.

None of these files contains a key.

7. Troubleshooting

Error messages from the API are shown as they are. Your key is removed from every error message and log line.