onnx-web/docs/chain-pipelines.md

Chain Pipelines

Chain pipelines are a new feature in v0.6 that allows you to run any combination of models on images of almost any size, by automatically splitting them into smaller tiles as needed. Individual models are run on each tile, then the results are recombined and passed on to the next stage.

Contents

• Chain Pipelines
  • Contents
  • Overview
    • Format
  • Stages
    • Blending Stages
      • Blend: Denoise
      • Blend: Grid
      • Blend: Img2img
      • Blend: Linear
      • Blend: Mask
    • Correction Stages
      • Correct: CodeFormer
      • Correct: GFPGAN
    • Compound Stages
      • Highres Stage
      • Upscale Stage
    • Persistence Stages
      • Persist: Disk
      • Persist: S3
    • Reduction Stages
      • Reduce: Crop
      • Reduce: Thumbnail
    • Source Stages
      • Source: Noise
      • Source: S3
      • Source: Txt2img
      • Source: URL
    • Upscaling Stages
      • Upscale: BSRGAN
      • Upscale: Highres
      • Upscale: Outpaint
      • Upscale: Real ESRGAN
      • Upscale: Simple
      • Upscale: Stable Diffusion
      • Upscale: SwinIR

Overview

Format

The /api/chain endpoint accepts a chain pipeline in JSON format and adds it to the queue of background jobs.

Pipelines are defined mostly through their stages, where each stage specifies a function to be run and the parameters for that function, including the name of the model to be used.

The output of the pipeline will not automatically be saved to disk, which is the case for the single-stage endpoints. You must use at least one persist-* stage. Persist stages can be placed anywhere in the pipeline and can also save intermediate output, such as the result of a source-txt2img stage before upscaling it.

{
  "stages": [
    {
      "name": "start",
      "type": "source-txt2img",
      "params": {
        "prompt": "a magical wizard"
      }
    },
    {
      "name": "expand",
      "type": "upscale-outpaint",
      "params": {
        "border": 256,
        "prompt": "a magical wizard in a robe fighting a dragon"
      }
    },
    {
      "name": "save-local",
      "type": "persist-disk",
      "params": {
        "tiles": "hd8k"
      }
    }
  ]
}

The complete schema can be found in api/schema.yaml and some example pipelines are available in common/pipelines.

Stages

Blending Stages

Blend: Denoise

Run fast non-local means denoising using cv2.

Blend: Grid

Combine the source images into a grid.

Blend: Img2img

Run an img2img pipeline.

Blend: Linear

Blend two images using linear interpolation (0.0 is the first image, 1.0 is the second).

Blend: Mask

Blend two images using a mask.

Correction Stages

Correct: CodeFormer

Run correction using CodeFormer.

Correct: GFPGAN

Run correction using GFPGAN.

Compound Stages

Not currently available through JSON API.

Highres Stage

Prep one or more highres iterations. Each iteration is an upscale stage followed by img2img.

Upscale Stage

Prep upscale and correction stages.

Persistence Stages

Persist: Disk

Save all of the sources to disk.

Persist: S3

Save all of the sources to an S3 bucket.

Reduction Stages

Reduce: Crop

Crop a section out of each source.

Reduce: Thumbnail

Downscale each image into a thumbnail of itself.

Source Stages

Source: Noise

Create a new source using a noise generator.

Source: S3

Load a new source from an S3 bucket.

Source: Txt2img

Run a txt2img pipeline.

Source: URL

Load a new source from a URL.

Upscaling Stages

Upscale: BSRGAN

Upscaling stage using BSRGAN.

Upscale: Highres

Upscaling stage using highres.

Upscale: Outpaint

Upscaling stage using outpainting. This adds empty borders to the source image, optionally fills them with noise, and then runs inpainting on those areas.

Upscale: Real ESRGAN

Upscaling stage using the Real ESRGAN upscaling models, available in x2 and x4 versions:

• https://github.com/xinntao/Real-ESRGAN/releases

Upscale: Simple

Upscaling stage using bilinear or Lanczos upscaling.

Upscale: Stable Diffusion

Upscaling stage using the Stable Diffusion x4 upscaling model:

• https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler
• https://huggingface.co/ssube/stable-diffusion-x4-upscaler-onnx

Upscale: SwinIR

Upscaling stage using SwinIR.