Workflow:AUTOMATIC1111 Stable diffusion webui Image to image generation
| Knowledge Sources | |
|---|---|
| Domains | Image_Generation, Stable_Diffusion, Inpainting, Generative_AI |
| Last Updated | 2026-02-08 08:00 GMT |
Overview
End-to-end process for transforming existing images using Stable Diffusion with support for standard img2img, sketch input, inpainting, and batch processing modes.
Description
This workflow covers the image-to-image generation pipeline that takes an existing image as input and transforms it guided by a text prompt. Unlike txt2img which starts from pure noise, img2img encodes the input image into latent space and adds noise proportional to the denoising strength, then denoises back to produce a modified version. The workflow supports five distinct input modes: standard img2img, sketch-based generation, inpainting with a drawn mask, inpaint sketch with automatic change detection, and mask upload. A batch processing mode handles multiple images from a directory with optional per-image masks.
Usage
Execute this workflow when you have an existing image that you want to modify, enhance, or selectively edit using Stable Diffusion. Use inpainting mode when you need to regenerate specific regions of an image while preserving the rest. Use batch mode when processing multiple images with consistent settings.
Execution Steps
Step 1: Input image selection and mode configuration
Select the input mode from the five available options: standard img2img (upload a reference image), sketch (draw on a canvas), inpaint (upload image and draw a mask to specify regions to regenerate), inpaint sketch (the system detects changed pixels as the mask), or upload mask (provide a separate mask image). For batch processing, specify an input directory and optionally a mask directory. Each mode determines how the initial image and optional mask are prepared.
Key considerations:
- Standard img2img works best for global style transfer and variations
- Inpainting requires a binary mask where white areas are regenerated
- Batch mode can extract generation parameters from PNG metadata for per-image settings
- The resize mode controls how input images are scaled to match the target dimensions
Step 2: Prompt and parameter configuration
Configure the text prompt, negative prompt, and generation parameters as in txt2img. The critical additional parameter is denoising strength (0.0 to 1.0), which controls how much the output differs from the input. Lower values preserve more of the original image; higher values allow more creative freedom. For inpainting, configure the masked content fill mode (fill, original, latent noise, latent nothing) and mask blur radius.
Key considerations:
- Denoising strength of 0.0 returns the original image; 1.0 is equivalent to txt2img
- Mask blur softens the boundary between inpainted and preserved regions
- Inpaint at full resolution crops and processes only the masked region for higher detail
- The padding parameter controls how much context around the masked region is included
Step 3: Image preprocessing and latent encoding
Resize the input image to the target dimensions using the selected resize mode (just resize, crop and resize, resize and fill, or latent upscale). For inpainting, process the mask by applying blur, creating a binary threshold, and optionally inverting it. Encode the preprocessed image through the VAE encoder into latent space. For batch mode, iterate through the input directory loading each image and its corresponding mask.
Key considerations:
- Crop and resize preserves aspect ratio by center-cropping
- Resize and fill preserves aspect ratio by padding with blurred content
- The latent representation is 8x smaller in each spatial dimension than the pixel image
- Color correction can be applied to match the output color distribution to the input
Step 4: Noise addition and guided denoising
Add noise to the encoded latent based on the denoising strength and sampling schedule. The amount of noise determines the starting point on the diffusion timeline: lower denoising strength starts closer to the clean image, higher starts closer to pure noise. Execute the sampling loop with CFG guidance from the encoded prompts. For inpainting, the denoiser blends the noised original latent with the denoised prediction at each step according to the mask, preserving unmasked regions.
Key considerations:
- The effective number of steps is reduced proportionally to the denoising strength
- Masked regions receive full denoising while unmasked areas maintain the original content
- Soft inpainting (optional extension) provides gradual blending at mask boundaries
- The same sampler and scheduler options from txt2img are available
Step 5: VAE decoding and output composition
Decode the final denoised latent through the VAE decoder to produce the output pixel image. For inpainting, composite the generated content into the original image using the mask, with optional overlay blending at the boundaries. Apply optional post-processing (face restoration, color correction). Save the output with generation metadata embedded in the PNG. For batch processing, repeat the entire process for each input image.
Key considerations:
- The overlay step ensures seamless blending between inpainted and original regions
- Face restoration can be applied selectively to the output
- Batch processing preserves individual image metadata and supports custom output directories
- Generation info includes the denoising strength and mask parameters for reproducibility