Search
English
Contents
  1. ComfyUI Tutorial
Getting started
  1. What is ComfyUI
  2. Install ComfyUI
  3. Interface tour
  4. Model downloads and folders
  5. First image (text to image)
Core techniques
  1. KSampler parameters
  2. Prompting techniques
  3. Image to image (img2img)
  4. Inpainting and outpainting
  5. LoRA
  6. ControlNet
  7. Image upscaling
Advanced Topics
  1. Custom Nodes and Manager
  2. Flux and other new-generation models
  3. AI video generation
  4. Automate with the API
  5. FAQ and troubleshooting

ComfyUI Image-to-Image Tutorial: Generate New Images from References

Image to image (img2img) starts with an existing image instead of pure noise. You can use it to:

  • Turn a hand-drawn sketch into a finished illustration
  • Transform a photo into a watercolor, anime, or other style
  • Keep the composition while changing only the details or colors

How it differs from text to image

Text to image starts with Empty Latent Image (an empty latent); image to image converts an image to a latent with VAE Encode, then passes it to KSampler:

Text to image: [Empty Latent Image] ──────────────────────▶ [KSampler denoise=1.0]
Image to image: [Load Image] ─▶ [VAE Encode] ─LATENT────────▶ [KSampler denoise=0.5]

Build the workflow

Modify the text-to-image workflow:

  1. Delete Empty Latent Image.
  2. Add Load Image, then click to upload an image (or drag an image directly onto the node).
  3. Add VAE Encode:
    • Connect pixels to the IMAGE output of Load Image.
    • Connect vae to the VAE output of Load Checkpoint.
  4. Connect the LATENT output of VAE Encode to latent_image on KSampler.
  5. Set denoise on KSampler to about 0.5.
ComfyUI image-to-image workflow
Image to image: Load Image goes through VAE Encode to KSampler, with denoise set to 0.6

Templates often include ready-to-use image-to-image workflows.

denoise is key

denoiseEffect
0.2–0.35Almost no change; only details and texture are refined
0.4–0.6Composition and colors are preserved, while the content and style change noticeably
0.7–0.85Only the general composition and color blocks are preserved
1.0Completely ignores the original image, equivalent to text to image

Start at 0.5, then adjust up or down based on the result.

Image dimensions

VAE Encode uses the original image dimensions. An image that is too large can be very slow or even run out of memory; one that is too small lacks detail. For best results, use dimensions close to the model's recommended resolution (for example, about 1024 pixels for SDXL).

Add a scaling node between Load Image and VAE Encode:

  • Upscale Image By: Scale by a factor, such as 0.5
  • Upscale Image: Set the width and height, with an option to choose how to crop
  • ImageScaleToTotalPixels: Scale to a specified total pixel count (for example, 1 megapixel) while preserving the aspect ratio

Ideally, the width and height should be multiples of 8 (most models recommend multiples of 64).

Examples

Sketch to finished artwork

  • Input: A sketch made with simple color blocks
  • Prompt: A complete description of the scene and style you want
  • denoise: 0.6–0.75 (to change the texture significantly while preserving the composition)

Photo to illustration style

  • Prompt: watercolor illustration of ... (describe the photo's subject and the target style)
  • denoise: 0.45–0.6
  • A style LoRA can improve the result. See LoRA.

Refine an AI-generated image

Generate an image, then run it again with a low denoise value (0.3) and the same prompt to add detail and fix minor flaws. This is also a common technique in the upscaling workflow.

Limitations of image to image

Image to image changes the entire image. To edit only one area (for example, change someone's clothes or remove a passerby), use Inpainting. To preserve the composition or pose while changing the content significantly, use ControlNet.

Updated October 12, 2026