Platform For AI:Deploy a ComfyUI service for AI video generation

Key limitations and considerations

Read this section before you deploy to avoid wasted resources and common mistakes.

• Deployment resources: We recommend GU30, NVIDIA A10, or NVIDIA T4 GPU instance types. The ml.gu7i.c16m60.1-gu30 instance type offers the best price-performance ratio.

  Important

  Each PAI-EAS instance runs a single ComfyUI process and supports only one GPU (single-instance single-GPU or multi-instance single-GPU). Multi-GPU parallel inference is not supported. Do not select multi-GPU instance types (such as 2×A10). Multi-GPU instance types waste resources and do not improve single-task performance.

• Horizontal scaling: To increase concurrency, use the API Edition and increase the number of replicas instead of selecting a multi-GPU instance type.

Choose a deployment edition

Select the edition that best fits your use case:

For more billing information, see Elastic Algorithm Service (EAS) billing.

Open the web interface

Click the name of the target service to open the Overview page, and then click Web Application in the upper-right corner.

If the page takes a long time to load, see Page freezes or takes too long to load .

Use a template workflow

ComfyUI provides prebuilt templates for common tasks.

1. Select a template (for example, Wan VACE Text to Video).

   Note

   Templates vary across ComfyUI image versions. If the sample template is not available, select a different template or load a workflow file from the local file system.

2. After the workflow loads, you may see errors about missing models. You can dismiss these errors (we recommend selecting the option to not show this message again).

3. Because model paths may have changed, running the workflow directly may produce the following error:

   Prompt outputs failed validation:
   UNETLoader:
   - Value not in list: unet_name: 'wan2.1_vace_14B_fp16.safetensors' not in (list of length 200)
   LoraLoader:
   - Value not in list: lora_name: 'Wan21_CausVid_14B_T2V_lora_rank32.safetensors' not in (list of length 364)

   In the Load models here section, reselect wan2.1_vace_14B_fp16.safetensors in the UNET loader and Wan21_CausVid_14B_T2V_lora_rank32.safetensors in the LoRA loader.

4. After the workflow runs successfully, the generated video appears in the Save Video section.

Use custom models and nodes

1. Make sure the service has a storage mount configured. After the service is deployed, the system automatically creates the following directory structure in the mounted OSS or NAS storage.

   data-oss/
   ├── custom_nodes/     # Custom node files (ComfyUI plugins)
   ├── models/           # Model files
   ├── input/
   ├── output/
   ├── unet/
   └── temp/

2. Upload model or node files. For example, you can use OSS and upload files to OSS through the console. For large files, see How do I upload large objects to OSS?.

   • Upload model files: Check the documentation of the source repository for the node that uses the model, and upload the model to the corresponding subdirectory under models. For example:

     • Checkpoint loader: Upload models to models/checkpoints.

     • Style model loader: Upload models to models/styles.

   • Upload node files: Upload custom nodes to the custom_nodes directory in the mounted storage.

3. Load new content.

   • Models: At the top of the page, click PaiCustom > Load New Models. If the model still does not appear, click PaiCustom > Restart Process. After the process restarts, refresh the browser page.

   • Nodes: Click Restart Process. After the process restarts, refresh the browser page.

Prepare the API request body

The request body format depends on the call method:

• Synchronous call: Wrap the workflow JSON content inside a prompt key-value pair.

• Asynchronous call: The request body is the workflow JSON content directly.

Because the Wan VACE Text to Video workflow above is time-consuming, the following test workflow is provided for easier testing (each run takes approximately 3 minutes).

{
    "prompt": {
        "3": {
            "inputs": {
                "seed": 423988542100860,
                "steps": 40,
                "cfg": 7,
                "sampler_name": "dpmpp_sde_gpu",
                "scheduler": "karras",
                "denoise": 1,
                "model": [
                    "4",
                    0
                ],
                "positive": [
                    "6",
                    0
                ],
                "negative": [
                    "7",
                    0
                ],
                "latent_image": [
                    "5",
                    0
                ]
            },
            "class_type": "KSampler",
            "_meta": {
                "title": "KSampler"
            }
        },
        "4": {
            "inputs": {
                "ckpt_name": "LandscapeBING_v10.safetensors"
            },
            "class_type": "CheckpointLoaderSimple",
            "_meta": {
                "title": "Checkpoint Loader (Simple)"
            }
        },
        "5": {
            "inputs": {
                "width": 720,
                "height": 1280,
                "batch_size": 1
            },
            "class_type": "EmptyLatentImage",
            "_meta": {
                "title": "Empty Latent"
            }
        },
        "6": {
            "inputs": {
                "text": "Rocket takes off from the ground, fire，sky, airplane",
                "speak_and_recognation": {
                    "__value__": [
                        false,
                        true
                    ]
                },
                "clip": [
                    "4",
                    1
                ]
            },
            "class_type": "CLIPTextEncode",
            "_meta": {
                "title": "CLIP Text Encoder"
            }
        },
        "7": {
            "inputs": {
                "text": "",
                "speak_and_recognation": {
                    "__value__": [
                        false,
                        true
                    ]
                },
                "clip": [
                    "4",
                    1
                ]
            },
            "class_type": "CLIPTextEncode",
            "_meta": {
                "title": "CLIP Text Encoder"
            }
        },
        "8": {
            "inputs": {
                "samples": [
                    "3",
                    0
                ],
                "vae": [
                    "4",
                    2
                ]
            },
            "class_type": "VAEDecode",
            "_meta": {
                "title": "VAE Decode"
            }
        },
        "9": {
            "inputs": {
                "filename_prefix": "ComfyUI",
                "images": [
                    "8",
                    0
                ]
            },
            "class_type": "SaveImage",
            "_meta": {
                "title": "Save Image"
            }
        },
        "13": {
            "inputs": {
                "seed": 788620942678235,
                "steps": 40,
                "cfg": 2.5,
                "sampler_name": "euler_ancestral",
                "scheduler": "karras",
                "denoise": 1,
                "model": [
                    "17",
                    0
                ],
                "positive": [
                    "16",
                    0
                ],
                "negative": [
                    "16",
                    1
                ],
                "latent_image": [
                    "16",
                    2
                ]
            },
            "class_type": "KSampler",
            "_meta": {
                "title": "KSampler"
            }
        },
        "14": {
            "inputs": {
                "samples": [
                    "13",
                    0
                ],
                "vae": [
                    "18",
                    2
                ]
            },
            "class_type": "VAEDecode",
            "_meta": {
                "title": "VAE Decode"
            }
        },
        "15": {
            "inputs": {
                "filename_prefix": "ComfyUI",
                "fps": 10.000000000000002,
                "lossless": false,
                "quality": 85,
                "method": "default",
                "images": [
                    "14",
                    0
                ]
            },
            "class_type": "SaveAnimatedWEBP",
            "_meta": {
                "title": "Save WEBP"
            }
        },
        "16": {
            "inputs": {
                "width": 512,
                "height": 768,
                "video_frames": 35,
                "motion_bucket_id": 140,
                "fps": 15,
                "augmentation_level": 0.15000000000000002,
                "clip_vision": [
                    "18",
                    1
                ],
                "init_image": [
                    "8",
                    0
                ],
                "vae": [
                    "18",
                    2
                ]
            },
            "class_type": "SVD_img2vid_Conditioning",
            "_meta": {
                "title": "SVD_Image to Video_Conditioning"
            }
        },
        "17": {
            "inputs": {
                "min_cfg": 1,
                "model": [
                    "18",
                    0
                ]
            },
            "class_type": "VideoLinearCFGGuidance",
            "_meta": {
                "title": "Linear CFG Guidance"
            }
        },
        "18": {
            "inputs": {
                "ckpt_name": "svd_xt.safetensors"
            },
            "class_type": "ImageOnlyCheckpointLoader",
            "_meta": {
                "title": "Checkpoint Loader (Image Only)"
            }
        },
        "19": {
            "inputs": {
                "frame_rate": 10,
                "loop_count": 0,
                "filename_prefix": "comfyUI",
                "format": "video/h264-mp4",
                "pix_fmt": "yuv420p",
                "crf": 20,
                "save_metadata": true,
                "trim_to_audio": false,
                "pingpong": false,
                "save_output": true,
                "images": [
                    "14",
                    0
                ]
            },
            "class_type": "VHS_VideoCombine",
            "_meta": {
                "title": "Combine to Video"
            }
        }
    }
}

{
    "3": {
        "inputs": {
            "seed": 423988542100860,
            "steps": 40,
            "cfg": 7,
            "sampler_name": "dpmpp_sde_gpu",
            "scheduler": "karras",
            "denoise": 1,
            "model": [
                "4",
                0
            ],
            "positive": [
                "6",
                0
            ],
            "negative": [
                "7",
                0
            ],
            "latent_image": [
                "5",
                0
            ]
        },
        "class_type": "KSampler",
        "_meta": {
            "title": "KSampler"
        }
    },
    "4": {
        "inputs": {
            "ckpt_name": "LandscapeBING_v10.safetensors"
        },
        "class_type": "CheckpointLoaderSimple",
        "_meta": {
            "title": "Checkpoint Loader (Simple)"
        }
    },
    "5": {
        "inputs": {
            "width": 720,
            "height": 1280,
            "batch_size": 1
        },
        "class_type": "EmptyLatentImage",
        "_meta": {
            "title": "Empty Latent"
        }
    },
    "6": {
        "inputs": {
            "text": "Rocket takes off from the ground, fire，sky, airplane",
            "speak_and_recognation": {
                "__value__": [
                    false,
                    true
                ]
            },
            "clip": [
                "4",
                1
            ]
        },
        "class_type": "CLIPTextEncode",
        "_meta": {
            "title": "CLIP Text Encoder"
        }
    },
    "7": {
        "inputs": {
            "text": "",
            "speak_and_recognation": {
                "__value__": [
                    false,
                    true
                ]
            },
            "clip": [
                "4",
                1
            ]
        },
        "class_type": "CLIPTextEncode",
        "_meta": {
            "title": "CLIP Text Encoder"
        }
    },
    "8": {
        "inputs": {
            "samples": [
                "3",
                0
            ],
            "vae": [
                "4",
                2
            ]
        },
        "class_type": "VAEDecode",
        "_meta": {
            "title": "VAE Decode"
        }
    },
    "9": {
        "inputs": {
            "filename_prefix": "ComfyUI",
            "images": [
                "8",
                0
            ]
        },
        "class_type": "SaveImage",
        "_meta": {
            "title": "Save Image"
        }
    },
    "13": {
        "inputs": {
            "seed": 788620942678235,
            "steps": 40,
            "cfg": 2.5,
            "sampler_name": "euler_ancestral",
            "scheduler": "karras",
            "denoise": 1,
            "model": [
                "17",
                0
            ],
            "positive": [
                "16",
                0
            ],
            "negative": [
                "16",
                1
            ],
            "latent_image": [
                "16",
                2
            ]
        },
        "class_type": "KSampler",
        "_meta": {
            "title": "KSampler"
        }
    },
    "14": {
        "inputs": {
            "samples": [
                "13",
                0
            ],
            "vae": [
                "18",
                2
            ]
        },
        "class_type": "VAEDecode",
        "_meta": {
            "title": "VAE Decode"
        }
    },
    "15": {
        "inputs": {
            "filename_prefix": "ComfyUI",
            "fps": 10.000000000000002,
            "lossless": false,
            "quality": 85,
            "method": "default",
            "images": [
                "14",
                0
            ]
        },
        "class_type": "SaveAnimatedWEBP",
        "_meta": {
            "title": "Save WEBP"
        }
    },
    "16": {
        "inputs": {
            "width": 512,
            "height": 768,
            "video_frames": 35,
            "motion_bucket_id": 140,
            "fps": 15,
            "augmentation_level": 0.15000000000000002,
            "clip_vision": [
                "18",
                1
            ],
            "init_image": [
                "8",
                0
            ],
            "vae": [
                "18",
                2
            ]
        },
        "class_type": "SVD_img2vid_Conditioning",
        "_meta": {
            "title": "SVD_Image to Video_Conditioning"
        }
    },
    "17": {
        "inputs": {
            "min_cfg": 1,
            "model": [
                "18",
                0
            ]
        },
        "class_type": "VideoLinearCFGGuidance",
        "_meta": {
            "title": "Linear CFG Guidance"
        }
    },
    "18": {
        "inputs": {
            "ckpt_name": "svd_xt.safetensors"
        },
        "class_type": "ImageOnlyCheckpointLoader",
        "_meta": {
            "title": "Checkpoint Loader (Image Only)"
        }
    },
    "19": {
        "inputs": {
            "frame_rate": 10,
            "loop_count": 0,
            "filename_prefix": "comfyUI",
            "format": "video/h264-mp4",
            "pix_fmt": "yuv420p",
            "crf": 20,
            "save_metadata": true,
            "trim_to_audio": false,
            "pingpong": false,
            "save_output": true,
            "images": [
                "14",
                0
            ]
        },
        "class_type": "VHS_VideoCombine",
        "_meta": {
            "title": "Combine to Video"
        }
    }
}

Synchronous calls

Synchronous calls send requests directly to the PAI-EAS service instance. This method is suitable for development, testing, and low-concurrency scenarios.

Asynchronous calls

Asynchronous calls are designed for high-concurrency production workloads. This method uses the PAI-EAS queue service for load balancing. Requests return immediately, and you retrieve results by subscribing to the result queue.

1. View the call information.

   On the Inference Services tab, click the name of the target service to open the Overview page. In the Basic Information section, click View Call Information. On the Asynchronous Call tab of the Call Information dialog box, view the service access endpoint and token.

   

   The following sections use <EAS_ENDPOINT> to refer to the Internet Input Endpoint (if the caller is in the same VPC as PAI-EAS, you can use the VPC Input Endpoint), and <EAS_TOKEN> to refer to the Token.

2. Send a request.

   Code example:

   import requests
   import json
   
   service_url = "<EAS_ENDPOINT>"
   token = "<EAS_TOKEN>"
   
   if service_url[-1] == "/":
       service_url = service_url[:-1]
   
   session = requests.session()
   session.headers.update({"Authorization":token})
   
   # Request body - workflow JSON, no need to wrap with "prompt".
   # Use triple quotes to define it as a multi-line string, otherwise you need to capitalize the boolean values (true and false) in the workflow JSON.
   payload = """{
       '3':
       ...omitted
     }
     """
   payload = json.loads(payload)
   
   for i in range(5):
     # task_id is the key identifier for matching requests and results. Assign a unique value to each request to correlate with the queue results later.
     response = session.post(url=f'{service_url}/api_prompt?task_id=txt2img_{i}', json=payload)
     if response.status_code != 200:
       exit(f"send request error:{response.content}")
     else:
       print(f"send {i} success, index is {response.content}")

3. Subscribe to results.

   1. Install the eas_prediction SDK.

      pip install eas_prediction  --user

   2. Run the following code to retrieve the result.

      from eas_prediction import QueueClient
      from urllib.parse import urlparse, urlunparse
      
      service_url     = "<EAS_ENDPOINT>"
      token           = "<EAS_TOKEN>"
      
      # Parse service URL
      def parse_service_url(service_url):
          parsed = urlparse(service_url)
          service_domain = f"{parsed.scheme}://{parsed.netloc}"
          path_parts = [p for p in parsed.path.strip('/').split('/') if p]
          service_name = path_parts[-1]
          return service_domain, service_name
      
      service_domain, service_name = parse_service_url(service_url)
      print(f"service_domain: {service_domain}, service_name: {service_name}.")
      
      # Create result queue client
      sink_queue = QueueClient(service_domain, f'{service_name}/sink')
      sink_queue.set_token(token)
      sink_queue.init()
      
      # Listen for results
      watcher = sink_queue.watch(0, 5, auto_commit=False)
      for x in watcher.run():
          if 'task_id' in x.tags:
              print('index {} task_id is {}'.format(x.index, x.tags['task_id']))
          print(f'index {x.index} data is {x.data}')
          # Commit acknowledgment
          sink_queue.commit(x.index)

      Example response:

      index 42 task_id is txt2img_0
      index 42 data is b'[{"type": "executed", "data": {"node": "9", "output": {"images": [{"filename": "ComfyUI_1712647318_8e7f3c93-d2a8-4377-92d5-8eb552adc172_.png", "subfolder": "", "type": "output"}]}, "prompt_id": "c3c983b6-f92b-4dd5-b4dc-442db4d1736f"}}, {"type": "executed", "data": {"node": "15", "output": {"images": [{"filename": "ComfyUI_1712647895_.webp", "subfolder": "", "type": "output"}], "animated": [true]}, "prompt_id": "c3c983b6-f92b-4dd5-b4dc-442db4d1736f"}}, {"type": "executed", "data": {"node": "19", "output": {"gifs": [{"filename": "comfyUI_00001.mp4", "subfolder": "", "type": "output", "format": "video/h264-mp4"}]}, "prompt_id": "c3c983b6-f92b-4dd5-b4dc-442db4d1736f"}}, {"9": {"images": [{"filename": "ComfyUI_1712647318_8e7f3c93-d2a8-4377-92d5-8eb552adc172_.png", "subfolder": "", "type": "output"}]}, "15": {"images": [{"filename": "ComfyUI_1712647895_.webp", "subfolder": "", "type": "output"}], "animated": [true]}, "19": {"gifs": [{"filename": "comfyUI_00001.mp4", "subfolder": "", "type": "output", "format": "video/h264-mp4"}]}}]'
      index 43 task_id is txt2img_1
      index 43 data is b'[{"9": {"images": [{"filename": "ComfyUI_1712647318_8e7f3c93-d2a8-4377-92d5-8eb552adc172_.png", "subfolder": "", "type": "output"}]}, "15": {"images": [{"filename": "ComfyUI_1712647895_.webp", "subfolder": "", "type": "output"}], "animated": [true]}, "19": {"gifs": [{"filename": "comfyUI_00001.mp4", "subfolder": "", "type": "output", "format": "video/h264-mp4"}]}}]'
      index 44 task_id is txt2img_2
      index 44 data is b'[{"9": {"images": [{"filename": "ComfyUI_1712647318_8e7f3c93-d2a8-4377-92d5-8eb552adc172_.png", "subfolder": "", "type": "output"}]}, "15": {"images": [{"filename": "ComfyUI_1712647895_.webp", "subfolder": "", "type": "output"}], "animated": [true]}, "19": {"gifs": [{"filename": "comfyUI_00001.mp4", "subfolder": "", "type": "output", "format": "video/h264-mp4"}]}}]'
      index 45 task_id is txt2img_3
      index 45 data is b'[{"9": {"images": [{"filename": "ComfyUI_1712647318_8e7f3c93-d2a8-4377-92d5-8eb552adc172_.png", "subfolder": "", "type": "output"}]}, "15": {"images": [{"filename": "ComfyUI_1712647895_.webp", "subfolder": "", "type": "output"}], "animated": [true]}, "19": {"gifs": [{"filename": "comfyUI_00001.mp4", "subfolder": "", "type": "output", "format": "video/h264-mp4"}]}}]'
      index 46 task_id is txt2img_4
      index 46 data is b'[{"9": {"images": [{"filename": "ComfyUI_1712647318_8e7f3c93-d2a8-4377-92d5-8eb552adc172_.png", "subfolder": "", "type": "output"}]}, "15": {"images": [{"filename": "ComfyUI_1712647895_.webp", "subfolder": "", "type": "output"}], "animated": [true]}, "19": {"gifs": [{"filename": "comfyUI_00001.mp4", "subfolder": "", "type": "output", "format": "video/h264-mp4"}]}}]'

      You can find the inference result files in the output directory of the mounted storage.

Deployment issues

Model and node issues

Runtime errors

Other

Install Python packages

Update the image version

If the service has OSS or NAS storage mounted, custom models are stored in the OSS or NAS storage and are not affected by an image version update.

1. On the service details page, click Update in the upper-right corner.

2. If the service was deployed through scenario-based model deployment, switch to custom model deployment.

3. In the Service Configuration section, edit the JSON configuration and update the image field under containers to the desired version (for example, change 1.9 to the target version).

4. Click Update Directly.

Generate a sign-in-free URL with a custom validity period

You can use an API to generate a sign-in-free web access URL with a specified validity period.

1. Go to the DescribeServiceSignedUrl API page.

2. Select the service endpoint.

3. Set the following parameters:

   • ClusterId (Region) and ServiceName: Enter the region and name of the service.

     On the EAS service overview page, the Service Name field shows the ServiceName value. The region is embedded in the node image address (for example, cn-hangzhou corresponds to the ClusterId region value). You can click View Web Application in the upper-right corner to open the service.

   • Type (Page type): Select webview from the drop-down list.

   • Expire (Expiration time): Enter an integer in seconds. The maximum value is 43200 seconds (12 hours). Set this as needed.

   • Internal (VPC access): Set to false (public access) or true (VPC access).

4. Click Call API. The SignedUrl field in the response contains the sign-in-free web access URL.

Change the default language of the WebUI

How Cluster Edition works

The following diagram shows the architecture:

Cluster Edition is designed for multi-user scenarios. It introduces a proxy layer that decouples clients from backend inference instances.

• Each user has an independent backend environment and working directory, but shares the backend GPU inference instance pool.

• When a user sends an inference request, the proxy selects an available instance from the pool to handle the request.

This time-multiplexing mechanism isolates user environments while maximizing GPU resource utilization and reducing inference costs in multi-user scenarios.