Glossary
img2img Workflow
A step-by-step guide to using image-to-image generation to restyle, refine, or repurpose existing photos and renders.
img2img (image-to-image) generation starts with an existing image rather than pure random noise. The pipeline adds a controlled amount of noise to your reference image, then runs the denoising process guided by your prompt, producing an output that blends your reference's structure with the prompt's direction.
The key parameter is denoising strength (also called noise strength or img2img strength): 0.0 returns the image unchanged; 1.0 is effectively pure text-to-image generation that merely uses the reference dimensions. The practical creative range is 0.3–0.75.
Common use cases: (1) Photo-to-illustration: upload a photo, prompt a painting style, set strength 0.6. (2) Render refinement: take a mediocre render, drop it into img2img at 0.35, and let the model 'polish' it while preserving composition. (3) Consistent character variation: generate a base image of your character, then img2img at 0.4 with different outfit prompts to maintain face identity.
Resolution considerations: img2img works best when input and output resolution match. Upscaling a low-res input before running img2img avoids blurriness artefacts in the result.
ControlNet + img2img: combining img2img with a ControlNet depth or pose map gives you both structural reference and creative freedom — the ControlNet locks macro composition while denoising strength controls detail variation.
Frequently Asked Questions
- What denoising strength should I use for subtle style changes?
- 0.25–0.45 for subtle changes that preserve most of the original composition. 0.55–0.75 for significant style shifts that keep only the rough layout.
- Can img2img improve image quality?
- Yes. Running a lower-quality render through img2img at 0.3–0.5 with a quality-focused prompt is a common post-processing trick to add detail without changing composition.
- What is the difference between img2img and inpainting?
- img2img applies changes to the whole image. Inpainting applies changes only to a masked region, leaving the rest untouched.
- Does img2img work for videos?
- Frame-by-frame img2img (with temporal consistency measures) is the basis of AI video style transfer, but dedicated video diffusion models like Kling or Stable Video Diffusion handle this more coherently.