ComfyUI API - A Stateless and Extendable API for ComfyUI

A simple wrapper that facilitates using ComfyUI as a stateless API, either by receiving images in the response, or by sending completed images to a webhook

• ComfyUI API - A Stateless and Extendable API for ComfyUI
  • Download and Usage
  • Features
  • Probes
  • API Configuration Guide
    • Environment Variables
    • Configuration Details
    • Additional Notes
  • Webhooks
    • output.complete
    • prompt.failed
  • System Events
    • status
    • progress
    • executing
    • execution_start
    • execution_cached
    • executed
    • execution_success
    • execution_interrupted
    • execution_error
  • Generating New Workflow Endpoints
    • Automating with Claude 3.5 Sonnet
  • Prebuilt Docker Images
  • Considerations for Running on SaladCloud
  • Contributing
  • Testing
    • Required Models
    • Running Tests
  • Architecture

Download and Usage

Either use a pre-built Docker image, or build your own.

Download the latest version from the release page, and copy it into your existing ComfyUI dockerfile. You can find good base dockerfiles in the docker directory. There are also example dockerfiles for popular models in the SaladCloud Recipes Repo.

If you have your own ComfyUI dockerfile, you can add the comfyui-api server to it like so:

# Change this to the version you want to use
ARG api_version=1.8.2


# Download the comfyui-api binary, and make it executable
ADD https://github.com/SaladTechnologies/comfyui-api/releases/download/${api_version}/comfyui-api .
RUN chmod +x comfyui-api

# Set CMD to launch the comfyui-api binary. The comfyui-api binary will launch ComfyUI as a child process.
CMD ["./comfyui-api"]

The server will be available on port 3000 by default, but this can be customized with the PORT environment variable.

The server hosts swagger docs at /docs, which can be used to interact with the API.

Features

• Full Power Of ComfyUI: The server supports the full ComfyUI /prompt API, and can be used to execute any ComfyUI workflow.
• Verified Model/Workflow Support: Stable Diffusion 1.5, Stable Diffusion XL, Stable Diffusion 3.5, Flux, AnimateDiff, LTX Video, Hunyuan Video, CogVideoX, Mochi Video, Cosmos 1.0. My assumption is more model types are supported, but these are the ones I have verified.
• Stateless API: The server is stateless, and can be scaled horizontally to handle more requests.
• Swagger Docs: The server hosts swagger docs at /docs, which can be used to interact with the API.
• "Synchronous" Support: The server will return base64-encoded images directly in the response, if no webhook is provided.
• Webhook Support: The server can send completed images to a webhook, which can be used to store images, or to send them to a user.
• Easily Submit Images: The server can accept images as base64-encoded strings, or as URLs to images. This makes image-to-image workflows much easier to use.
• Warmup Workflow: The server can be configured to run a warmup workflow on startup, which can be used to load and warm up models, and to ensure the server is ready to accept requests.
• Return Images In PNG (default), JPEG, or WebP: The server can return images in PNG, JPEG, or WebP format, via a parameter in the API request. Most options supported by sharp are supported.
• Probes: The server has two probes, /health and /ready, which can be used to check the server's health and readiness to receive traffic.
• Dynamic Workflow Endpoints: Automatically mount new workflow endpoints by adding conforming .js or .ts files to the /workflows directory in your docker image. See below for more information. A Claude 3.5 Sonnet prompt is included to assist in automating this process.
• Bring Your Own Models And Extensions: Use any model or extension you want by adding them to the normal ComfyUI directories /opt/ComfyUI/.
• Works Great with SaladCloud: The server is designed to work well with SaladCloud, and can be used to host ComfyUI on the SaladCloud platform. It is likely to work well with other platforms as well.
• Single Binary: The server is distributed as a single binary, and can be run with no dependencies.
• Websocket Events Via Webhook: The server can forward ComfyUI websocket events to a configured webhook, which can be used to monitor the progress of a workflow.
• Friendly License: The server is distributed under the MIT license, and can be used for any purpose. All of its dependencies are also MIT or Apache 2.0 licensed, except ComfyUI itself, which is GPL-3.0 licensed.

Probes

The server has two probes, /health and /ready.

• The /health probe will return a 200 status code once the warmup workflow has completed. It will stay healthy as long as the server is running, even if ComfyUI crashes.
• The /ready probe will also return a 200 status code once the warmup workflow has completed. It will return a 503 status code if ComfyUI is not running, such as in the case it has crashed, but is being automatically restarted. If you have set MAX_QUEUE_DEPTH to a non-zero value, it will return a 503 status code if ComfyUI's queue has reached the maximum depth.

Here's a markdown guide to configuring the application based on the provided config.ts file:

API Configuration Guide

This guide provides an overview of how to configure the application using environment variables.

Environment Variables

The following table lists the available environment variables and their default values. For historical reasons, the default values mostly assume this will run on top of an ai-dock image, but we currently provide our own more minimal image here in this repo.

Configuration Details

1. ComfyUI Settings:

   • The application uses the CMD environment variable to specify the command for launching ComfyUI.
   • ComfyUI is accessed at http://${DIRECT_ADDRESS}:${COMFYUI_PORT_HOST}.

2. Wrapper Settings:

   • The wrapper API listens on HOST:PORT.
   • It can be accessed at http://localhost:${PORT}.
   • Use an IPv6 address for HOST when deploying on Salad. This is the default behavior.

3. Startup Checks:

   • The application performs startup checks at intervals specified by STARTUP_CHECK_INTERVAL_S.
   • It will attempt up to STARTUP_CHECK_MAX_TRIES before giving up.

4. Directories:

   • The application uses the COMFY_HOME environment variable to locate the ComfyUI installation.
   • Output files are stored in OUTPUT_DIR.
   • Input files are read from INPUT_DIR.
   • Model files are located in MODEL_DIR.
   • Workflow files are stored in WORKFLOW_DIR. See below for more information.

5. Warmup Prompt:

   • If WARMUP_PROMPT_FILE is set, the application will load and parse a warmup prompt from this file.
   • The checkpoint used in this prompt can be used as the default for workflow models.

6. Models:

   • The application scans the MODEL_DIR for subdirectories and creates configurations for each model type found.
   • Each model type will have its directory path, list of available models, and a Zod enum for validation.
   • The model names are exposed via the GET /models endpoint, and via the config object throughout the application.

7. ComfyUI Description:

   • The application retrieves available samplers and schedulers from ComfyUI itself.
   • This information is used to create Zod enums for validation.

Additional Notes

• The application uses Zod for runtime type checking and validation of configuration values.
• The configuration includes setup for both the wrapper application and ComfyUI itself.

Remember to set these environment variables according to your specific deployment needs before running the application.

Webhooks

ComfyUI API sends two types of webhooks: System Events, which are emitted by ComfyUI itself, and Workflow Events, which are emitted by the API server. See System Events for more information on System Events.

If a user includes the .webhook field in a request to /prompt or any of the workflow endpoints, the server will send any completed outputs to the webhook URL provided in the request. It will also send a webhook if the request fails.

For successful requests, every output from the workflow will be sent as individual webhook requests. That means if your request generates 4 images, you will receive 4 webhook requests, each with a single image.

output.complete

The webhook event name for a completed output is output.complete. The webhook will have the following schema:

{
  "event": "output.complete",
  "image": "base64-encoded-image",
  "id": "request-id",
  "filename": "output-filename.png",
  "prompt": {}
}

prompt.failed

The webhook event name for a failed request is prompt.failed. The webhook will have the following schema:

{
  "event": "prompt.failed",
  "error": "error-message",
  "id": "request-id",
  "prompt": {}
}

System Events

ComfyUI emits a number of events over websocket during the course of a workflow. These can be configured to be sent to a webhook using the SYSTEM_WEBHOOK_URL and SYSTEM_WEBHOOK_EVENTS environment variables. Additionally, any environment variable starting with SYSTEM_META_ will be sent as metadata with the event.

All webhooks have the same format, which is as follows:

{
  "event": "event_name",
  "data": {},
  "metadata": {}
}

When running on SaladCloud, .metadata will always include salad_container_group_id and salad_machine_id.

The following events are available:

• "status"
• "progress"
• "executing"
• "execution_start"
• "execution_cached"
• "executed"
• "execution_success"
• "execution_interrupted"
• "execution_error"

The SYSTEM_WEBHOOK_EVENTS environment variable should be a comma-separated list of the events you want to send to the webhook. If not set, no events will be sent.

The event name received in the webhook will be comfy.${event_name}, i.e. comfy.progress.

Example:

export SYSTEM_WEBHOOK_EVENTS="progress,execution_start,execution_success,execution_error"

This will cause the API to send the progress, execution_start, execution_success, and execution_error events to the webhook.

The SYSTEM_META_* environment variables can be used to add metadata to the webhook events. For example:

export SYSTEM_META_batch=abc
export SYSTEM_META_purpose=testing

Will add {"batch": "abc", "purpose": "testing"} to the .metadata field on system webhooks.

The following are the schemas for the event data that will be sent to the webhook. This will populate the .data field on the webhook.

status

{
  "type": "status",
  "data": {
    "status": {
      "exec_info": {
        "queue_remaining": 3
      }
    }
  },
  "sid": "abc123"
}

progress

{
  "type": "progress",
  "data": {
    "value": 45,
    "max": 100,
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "node": "42"
  },
  "sid": "xyz789"
}

executing

{
  "type": "executing",
  "data": {
    "node": "42",
    "display_node": "42",
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000"
  },
  "sid": "xyz789"
}

execution_start

{
  "type": "execution_start",
  "data": {
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "timestamp": 1705505423000
  },
  "sid": "xyz789"
}

execution_cached

{
  "type": "execution_cached",
  "data": {
    "nodes": ["42", "7", "13"],
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "timestamp": 1705505423000
  },
  "sid": "xyz789"
}

executed

{
  "type": "executed",
  "data": {
    "node": "42",
    "display_node": "42",
    "output": {},
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000"
  },
  "sid": "xyz789"
}

execution_success

{
  "type": "execution_success",
  "data": {
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "timestamp": 1705505423000
  },
  "sid": "xyz789"
}

execution_interrupted

{
  "type": "execution_interrupted",
  "data": {
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "node_id": "42",
    "node_type": "KSampler",
    "executed": []
  },
  "sid": "xyz789"
}

execution_error

{
  "type": "execution_error",
  "data": {
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "node_id": "42",
    "node_type": "KSampler",
    "executed": [],
    "exception_message": "CUDA out of memory. Tried to allocate 2.20 GiB",
    "exception_type": "RuntimeError",
    "traceback": "Traceback (most recent call last):\n  File \"nodes.py\", line 245, in sample\n    samples = sampler.sample(model, noise, steps)",
    "current_inputs": {
      "seed": 42,
      "steps": 20,
      "cfg": 7.5,
      "sampler_name": "euler"
    },
    "current_outputs": []
  },
  "sid": "xyz789"
}

Generating New Workflow Endpoints

Since the ComfyUI prompt format is a little obtuse, it's common to wrap the workflow endpoints with a more user-friendly interface.

This can be done by adding conforming .js or .ts files to the /workflows directory in your dockerfile. You can see some examples in ./workflows. Typescript files will be automatically transpiled to javascript files, so you can use either.

Here is an example text-to-image workflow file.

import { z } from "zod";
import config from "../config";

const ComfyNodeSchema = z.object({
  inputs: z.any(),
  class_type: z.string(),
  _meta: z.any().optional(),
});

type ComfyNode = z.infer<typeof ComfyNodeSchema>;
type ComfyPrompt = Record<string, ComfyNode>;

interface Workflow {
  RequestSchema: z.ZodObject<any, any>;
  generateWorkflow: (input: any) => Promise<ComfyPrompt> | ComfyPrompt;
  description?: string;
  summary?: string;
}

// This defaults the checkpoint to whatever was used in the warmup workflow
let checkpoint: any = config.models.checkpoints.enum.optional();
if (config.warmupCkpt) {
  checkpoint = checkpoint.default(config.warmupCkpt);
}

const RequestSchema = z.object({
  prompt: z.string().describe("The positive prompt for image generation"),
  negative_prompt: z
    .string()
    .optional()
    .default("text, watermark")
    .describe("The negative prompt for image generation"),
  width: z
    .number()
    .int()
    .min(256)
    .max(2048)
    .optional()
    .default(512)
    .describe("Width of the generated image"),
  height: z
    .number()
    .int()
    .min(256)
    .max(2048)
    .optional()
    .default(512)
    .describe("Height of the generated image"),
  seed: z
    .number()
    .int()
    .optional()
    .default(() => Math.floor(Math.random() * 100000000000))
    .describe("Seed for random number generation"),
  steps: z
    .number()
    .int()
    .min(1)
    .max(100)
    .optional()
    .default(20)
    .describe("Number of sampling steps"),
  cfg_scale: z
    .number()
    .min(0)
    .max(20)
    .optional()
    .default(8)
    .describe("Classifier-free guidance scale"),
  sampler_name: config.samplers
    .optional()
    .default("euler")
    .describe("Name of the sampler to use"),
  scheduler: config.schedulers
    .optional()
    .default("normal")
    .describe("Type of scheduler to use"),
  denoise: z
    .number()
    .min(0)
    .max(1)
    .optional()
    .default(1)
    .describe("Denoising strength"),
  checkpoint,
});

type InputType = z.infer<typeof RequestSchema>;

function generateWorkflow(input: InputType): ComfyPrompt {
  return {
    "3": {
      inputs: {
        seed: input.seed,
        steps: input.steps,
        cfg: input.cfg_scale,
        sampler_name: input.sampler_name,
        scheduler: input.scheduler,
        denoise: input.denoise,
        model: ["4", 0],
        positive: ["6", 0],
        negative: ["7", 0],
        latent_image: ["5", 0],
      },
      class_type: "KSampler",
      _meta: {
        title: "KSampler",
      },
    },
    "4": {
      inputs: {
        ckpt_name: input.checkpoint,
      },
      class_type: "CheckpointLoaderSimple",
      _meta: {
        title: "Load Checkpoint",
      },
    },
    "5": {
      inputs: {
        width: input.width,
        height: input.height,
        batch_size: 1,
      },
      class_type: "EmptyLatentImage",
      _meta: {
        title: "Empty Latent Image",
      },
    },
    "6": {
      inputs: {
        text: input.prompt,
        clip: ["4", 1],
      },
      class_type: "CLIPTextEncode",
      _meta: {
        title: "CLIP Text Encode (Prompt)",
      },
    },
    "7": {
      inputs: {
        text: input.negative_prompt,
        clip: ["4", 1],
      },
      class_type: "CLIPTextEncode",
      _meta: {
        title: "CLIP Text Encode (Prompt)",
      },
    },
    "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",
      },
    },
  };
}

const workflow: Workflow = {
  RequestSchema,
  generateWorkflow,
  summary: "Text to Image",
  description: "Generate an image from a text prompt",
};

export default workflow;

Note your file MUST export a Workflow object, which contains a RequestSchema and a generateWorkflow function. The RequestSchema is a zod schema that describes the input to the workflow, and the generateWorkflow function takes the input and returns a ComfyUI API-format prompt.

The workflow endpoints will follow whatever directory structure you provide. For example, a directory structure like this:

/workflows
└── sdxl
    ├── img2img.ts
    ├── txt2img-with-refiner.ts
    └── txt2img.ts

Would yield the following endpoints:

• POST /workflows/sdxl/img2img
• POST /workflows/sdxl/txt2img-with-refiner
• POST /workflows/sdxl/txt2img

These endpoints will be present in the swagger docs, and can be used to interact with the API. If you provide descriptions in your zod schemas, these will be used to create a table of inputs in the swagger docs.

Automating with Claude 3.5 Sonnet

Note: This requires having an account with Anthropic, and your anthropic API key in the environment variable ANTHROPIC_API_KEY.

Creating these endpoints can be done mostly automatically by Claude 3.5 Sonnet, given the JSON prompt graph. A system prompt to do this is included here.

A script that uses this prompt to create endpoints is included here. It requires jq and curl to be installed.

./generate-workflow <inputFile> <outputFile>

Where <inputFile> is the JSON prompt graph, and <outputFile> is the output file to write the generated workflow to.

As with all AI-generated code, it is strongly recommended to review the generated code before using it in production.

Prebuilt Docker Images

You can find ready-to-go docker images under Packages in this repository.

The images are tagged with the comfyui-api version they are built with, and the comfyui version they are built for, along with their pytorch version and CUDA version. There are versions for both CUDA runtime and CUDA devel, so you can choose the one that best fits your needs.

The tag pattern is ghcr.io/saladtechnologies/comfyui-api:comfy<comfy-version>-api<api-version>-torch<pytorch-version>-cuda<cuda-version>-<runtime|devel> where:

• <comfy-version> is the version of ComfyUI used
• <api-version> is the version of the comfyui-api server
• <pytorch-version> is the version of PyTorch used
• <cuda-version> is the version of CUDA used
• <runtime|devel> is whether the image is built with the CUDA runtime or the CUDA devel image. The devel image is much larger, but includes the full CUDA toolkit, which is required for some custom nodes.

If the tag doesn't have api<api-version>, it does not include the api, and is just the ComfyUI base image.

Included in the API images are the following utilities:

• git
• curl
• wget
• unzip
• ComfyUI
• comfy cli

All of SaladCloud's image and video generation recipes are built on top of these images, so you can use them as a base for your own workflows. For examples of using this with custom models and nodes, check out the Salad Recipes repository on GitHub.

Considerations for Running on SaladCloud

• SaladCloud's Container Gateway has a 100s timeout. It is possible to construct very long running ComfyUI workflows, such as for video generation, that would exceed this timeout. In this scenario, you will need to either use a webhook to receive the results, or integrate with SaladCloud's Job Queues to handle long-running workflows.
• SaladCloud's maximum container image size is 35GB(compressed). The base comfyui-api image is around 3.25GB(compressed), so any models and extensions must fit in the remaining space.

Contributing

Contributions are welcome! ComfyUI is a powerful tool with MANY options, and it's likely that not all of them are currently supported by the comfyui-api server. Please open an issue with as much information as possible about the problem you're facing or the feature you need. If you have encountered a bug, please include the steps to reproduce it, and any relevant logs or error messages. If you are able, adding a failing test is the best way to ensure your issue is resolved quickly. Let's make productionizing ComfyUI as easy as possible!

Testing

Required Models

Automated tests for this project require model files to be present in the ./test/docker-image/models directory. The following models are required:

• AnimateLCM_sd15_t2v.ckpt - https://huggingface.co/wangfuyun/AnimateLCM/resolve/b78bbce/AnimateLCM_sd15_t2v.ckpt
• dreamshaper_8.safetensors - https://civitai.com/models/4384/dreamshaper
• flux1-schnell-fp8.safetensors - https://huggingface.co/Comfy-Org/flux1-schnell
• ltx-video-2b-v0.9.1.safetensors - https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.1.safetensors
• sd3.5_medium.safetensors - https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
• sd_xl_base_1.0.safetensors - https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
• sd_xl_refiner_1.0.safetensors - https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0
• clip_g.safetensors - https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_g.safetensors
• clip_l.safetensors - https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_l.safetensors
• t5xxl_fp16.safetensors - https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
• t5xxl_fp8_e4m3fn.safetensors - https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp8_e4m3fn_scaled.safetensors
• openpose-sd1.5-1.1.safetensors - https://huggingface.co/lllyasviel/control_v11p_sd15_openpose/resolve/main/diffusion_pytorch_model.fp16.safetensors
• hunyuan_video_t2v_720p_bf16.safetensors - https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged/tree/main/split_files/diffusion_models
• jump_V2.safetensors - https://civitai.com/models/193225?modelVersionId=235847
• llava_llama3_fp8_scaled.safetensors - https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged/tree/main/split_files/text_encoders
• hunyuan_video_vae_bf16.safetensors - https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged/tree/main/split_files/vae
• vae-ft-mse-840000-ema-pruned.ckpt - https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt
• THUDM/CogVideoX-2b - https://huggingface.co/THUDM/CogVideoX-2b
• mochi_preview_fp8_scaled.safetensors - https://huggingface.co/Comfy-Org/mochi_preview_repackaged/blob/main/all_in_one/mochi_preview_fp8_scaled.safetensors
• oldt5_xxl_fp8_e4m3fn_scaled.safetensors - https://huggingface.co/comfyanonymous/cosmos_1.0_text_encoder_and_VAE_ComfyUI/tree/main/text_encoders
• cosmos_cv8x8x8_1.0.safetensors - https://huggingface.co/comfyanonymous/cosmos_1.0_text_encoder_and_VAE_ComfyUI/blob/main/vae/cosmos_cv8x8x8_1.0.safetensors
• Cosmos-1_0-Diffusion-7B-Text2World.safetensors - https://huggingface.co/mcmonkey/cosmos-1.0/blob/main/Cosmos-1_0-Diffusion-7B-Text2World.safetensors

They should be in the correct comfyui directory structure, like so:

./test/docker-image/models
├── animatediff_models
│   └── AnimateLCM_sd15_t2v.ckpt
├── checkpoints
│   ├── dreamshaper_8.safetensors
│   ├── flux1-schnell-fp8.safetensors
│   ├── ltx-video-2b-v0.9.1.safetensors
|   ├── mochi_preview_fp8_scaled.safetensors
│   ├── sd3.5_medium.safetensors
│   ├── sd_xl_base_1.0.safetensors
│   └── sd_xl_refiner_1.0.safetensors
├── clip
│   ├── clip_g.safetensors
│   ├── clip_l.safetensors
│   ├── t5xxl_fp16.safetensors
│   └── t5xxl_fp8_e4m3fn.safetensors
├── CogVideo
│   └── CogVideo2B/
├── controlnet
│   ├── openpose-sd1.5-1.1.safetensors
├── diffusion_models
│   ├── hunyuan_video_t2v_720p_bf16.safetensors
|   └── Cosmos-1_0-Diffusion-7B-Text2World.safetensors
├── loras
│   ├── jump_V2.safetensors
├── text_encoders
│   ├── clip_l.safetensors
│   ├── llava_llama3_fp8_scaled.safetensors
|   └── oldt5_xxl_fp8_e4m3fn_scaled.safetensors
├── vae
│   ├── hunyuan_video_vae_bf16.safetensors
│   ├── vae-ft-mse-840000-ema-pruned.ckpt
│   └── cosmos_cv8x8x8_1.0.safetensors

Running Tests

In one terminal, start the test server:

docker compose up --build

--build is only needed the first time, or if you make changes to the server code

In another terminal, run the tests:

npm test

This will take quite a long time, and requires a minimum of 24gb of RAM. I did these tests on my RTX 3080ti Laptop Edition w/ 16gb VRAM, and 24gb WSL RAM. It takes about 30 minutes to run all the tests.

Architecture

The server is built with Fastify, a fast and low overhead web framework for Node.js. It sits in front of ComfyUI, and provides a RESTful API for interacting with ComfyUI.