What is img2img?
What happens when a diffusion model starts from a photo instead of noise, what the strength parameter controls, and how the mode relates to inpainting.
What does img2img do?
Image-to-image generation (img2img) is the mode in which a diffusion model starts from a noised copy of an existing picture instead of pure noise and denoises it under a text prompt. The 2021 SDEdit paper formalized the idea; in Stable Diffusion the default strength of 0.8 replaces most of the input while keeping its composition.
| Term | What it does | Where used | Limits |
|---|---|---|---|
| img2img | Re-renders the whole image from a partially noised copy | Style transfer, variations, sketch-to-render | Faces and text drift at high strength |
| Inpainting | Regenerates only a masked region | Retouching, object removal, undress apps | Fill is invented; seams at the mask edge |
| ControlNet-guided generation | Adds pose, depth or edge maps as conditioning | Pose-locked renders, consistent characters | Needs extra models; adds latency |
| Image-to-video | Animates a still image into a clip | Short video from one photo | Different model family; seconds of output |
How does img2img work?
Text-to-image starts from random noise and denoises for a fixed number of steps. img2img encodes the input image into latent space, adds noise up to a chosen level, and then runs only the remaining denoising steps.
With 50 steps and strength 0.8, the model skips the first 10 steps and runs 40, so the coarse layout of the input survives while textures, style and details are regenerated. The SDEdit paper shows that realism and faithfulness trade off along the noise level.
Given an input image with user guide of any type, SDEdit first adds noise to the input, then subsequently denoises the resulting image through the SDE prior to increase its realism.
— Meng et al., “SDEdit” (arXiv:2108.01073, 2021), accessed September 3, 2026
What does the strength parameter control?
Strength is the fraction of the diffusion process that is rerun. At 0.2–0.4 the output is a cleaned-up version of the input; at 0.5–0.7 style and materials change but the composition holds; at 0.8–1.0 only the rough silhouette remains and faces are redrawn.
Because every pixel can change, img2img on a photo of a person produces a new person unless a mask or a face-preserving adapter is added.
How do undress and nude generators use it?
Tools in the AI nude generator category use img2img for two jobs: turning a clothed reference into a stylized nude “inspired by” the input, and producing variations of a generated result.
For photoreal clothing removal they switch to inpainting, which keeps face and background intact; pure img2img at the strength needed to change clothing also changes the face. Most services combine both on Stable Diffusion checkpoints and add an upscaler.
How does img2img affect consent and identity?
The legal analysis follows the output, not the mode. If the result is an intimate image of an identifiable real person without consent, it is a digital forgery under the TAKE IT DOWN Act regardless of whether img2img, inpainting or a face swap produced it.
Even at high strength, pose, setting and clothing can still identify a person. Reviewed tools require consent confirmation for any uploaded reference under the rules in Responsible AI.
Consent Note
This site is a publisher: it reviews third-party tools and does not create, edit, accept or store images. Docs entries are general information, not legal advice.
Sources
- Meng et al., SDEdit (2021) — arXiv:2108.01073 via Hugging Face, accessed September 3, 2026.
- Diffusers docs: img2img pipeline — Hugging Face, accessed September 3, 2026.
- CompVis/stable-diffusion repository — GitHub, accessed September 3, 2026.
- Rombach et al., latent diffusion (2021) — arXiv:2112.10752 via Hugging Face, accessed September 3, 2026.
- TAKE IT DOWN Act, Public Law 119-12 — GovInfo, accessed September 3, 2026.
- How we rate — Clothoff AI.
FAQ
What is the difference between img2img and inpainting?
img2img re-renders the whole image from a noised copy, so everything can change; inpainting regenerates only a masked region and leaves the rest untouched. Undress apps use inpainting because they need the face and background preserved; img2img suits style transfer and variations.
What strength should I use?
It depends on the goal. Values of 0.2–0.4 clean up or restyle lightly, 0.5–0.7 change materials and style while keeping composition, and 0.8–1.0 keep only the rough layout. The Stable Diffusion reference default is 0.8; faces usually drift above 0.5.
Does img2img keep the person’s face?
Not reliably. At strengths above roughly 0.5 the face is redrawn and may no longer resemble the input. Tools that need identity preservation use inpainting with the face masked out, face-restoration models or reference adapters, which is why photoreal undress services rely on inpainting.
Where does the term come from?
From the Stable Diffusion command-line scripts released in August 2022, which shipped as txt2img.py and img2img.py. The method is older: the 2021 SDEdit paper described adding noise to an input image and denoising it with a diffusion prior. The name stuck across web interfaces and APIs.
Is img2img output copyrighted or original?
US copyright law protects human authorship; the Copyright Office has refused registration for images generated without sufficient human creative control, while accepting human-made selections and arrangements. An img2img render of someone else’s photo may also infringe that photo. This is general information, not legal advice.
Can img2img be used to remove clothing?
At high strength it produces a new body loosely following the input pose, with the face changed; that is a stylized nude, not a removal. Photoreal clothing removal uses inpainting. In either case an intimate image of an identifiable non-consenting person is a digital forgery under federal law.