Image-to-image AI makes a new picture from a picture you give it plus a written prompt. In the classic method, the AI blurs your image with digital noise, then cleans it up again while steering towards your prompt, so the result keeps your layout but changes the look. A strength (or denoise) setting decides how much of your original survives. Newer chat tools also let you upload a picture and simply say what to change.
- The idea: your picture is the starting point instead of blank noise. The prompt says what to turn it into.
- The dial: strength (called denoise in some apps) runs from 0 to 1. Low keeps your picture; high lets the AI wander.
- Three ways it shows up: start from your picture (classic img2img), borrow from a reference picture (style, face, product), or edit by instruction (“change only the background”).
- It always redraws. Even a light pass can shift small details, so check faces, hands and text.
- Use pictures you have the right to use, and never put a real person in a scene without their consent.
Imagine a chalk drawing left out in light rain. A short shower smudges it a little; a downpour leaves only faint shapes. Now an artist redraws over whatever is left, following your note (“make it a watercolour harbour”). Strength is how hard it rained. A light shower keeps almost everything; a storm keeps only the rough layout.
| Way | What your picture does | Use it for |
|---|---|---|
| Start from it (img2img) | Becomes the base: noise is added to it, then removed towards your prompt. | Use it for: sketch to finished image, restyling, cleaning up a nearly-right result. |
| Reference it | Steers the result: the AI borrows a style, face or product from it but draws a new picture. | Use it for: the same character or style across many images. |
| Edit by instruction | Gets changed in place: you say what to change and what to keep. | Use it for: a new background, an outfit, removing an object. |
What is image-to-image AI?
The plain point: it is text-to-image with a starting picture added.
Hugging Face, which publishes widely used open-source image software, defines it in one line: “Image-to-image is similar to text-to-image, but in addition to a prompt, you can also pass an initial image as a starting point for the diffusion process.”1 The short name is img2img; this site’s img2img meaning page gives the one-paragraph version.
The “diffusion process” is how most image models work: a diffusion model starts from random static and removes it step by step. Image-to-image begins from your picture instead.
Image-to-image keeps the bones of your picture and repaints the rest. How much it repaints is up to you.
How does image-to-image work, step by step?
The plain point: add noise to your picture, then let the model clean it up towards your prompt.
Hugging Face lays out four steps.1 In plain words:
- Shrink it into the model’s own format. Your image is turned into a compact set of numbers (its “latent” form), the internal format the model works in.
- Add noise. Random static is mixed in, like the rain on the chalk drawing.
- Clean it up towards the prompt. The model guesses which parts are noise and removes them, steered by your words.
- Turn it back into a picture. A decoder converts the numbers back into an image you can see.
The method has a name and a date. Researchers led by Chenlin Meng posted it in August 2021 as SDEdit, which “first adds noise to the input, then subsequently denoises the resulting image”.2 The original Stable Diffusion code shipped it as a script called img2img.py and credits SDEdit for the idea.3
This is why image-to-image never returns your exact picture. Every pixel passes through the noise and gets redrawn, so small details such as a logo or a fingertip can change even when the overall picture looks the same.
What does the strength (denoise) setting do?
The plain point: it sets how much noise goes in, and so how much of your picture comes back out.
Hugging Face describes both ends of the dial: “a higher strength value gives the model more ‘creativity’ to generate an image that’s different from the initial image; a strength value of 1.0 means the initial image is more or less ignored”, while a lower value keeps the result closer to the original.1
Strength also decides how many clean-up steps run. The same guide gives an example: with 50 steps and a strength of 0.8, the software adds 40 steps of noise and then removes it over 40 steps.1
Different apps name the dial differently:
- Hugging Face Diffusers calls it strength, from 0 to 1, starting at 0.8 unless you change it.4
- ComfyUI, the node-based image app, calls it denoise on its KSampler node. It runs from 0 to 1 and starts at 1.0, which ignores your image; lower values “preserve original structure”.5 See ComfyUI explained.
- Midjourney has no strength slider for this. It uses an image weight for image prompts, which defaults to 1 and runs from 0 to 3 in version 8.1.6 See the Midjourney cheat sheet.
Strength works with a second dial, guidance (often called CFG scale), which sets how strictly the model follows the prompt. Hugging Face suggests “a combination of low strength and low guidance_scale” for a result that resembles the original image without being tied tightly to the prompt.1 See CFG scale meaning. For the strength ranges this site works with, see how to use reference images; they are working advice, not vendor figures.
Start low and step up. A low strength fixes and polishes; a high one reimagines. At the very top, your picture barely matters.
Reference or edit: what is the difference?
The plain point: a reference steers a brand-new picture; an edit changes your picture and should leave the rest alone.
A reference picture is not the starting point. The AI draws from scratch but borrows something from your image: its style, a face, a product. Midjourney’s style references, for example, are described as “similar to image prompts but focuses solely on the style.”7 In open-source tools the usual add-on is IP-Adapter, which its authors describe as a way “to achieve image prompt capability” for text-to-image models.8 See style reference meaning.
An instruction edit is newer. You upload a picture and write the change in plain words. A 2022 research paper, InstructPix2Pix, set out “a method for editing images from human instructions”.9 Chat tools now work this way. Google’s guide for its Nano Banana image models says: “Provide an image and describe your change. The model will match the original image’s style, lighting, and perspective.”10 OpenAI’s help page says: “To edit an image, select an image generated by ChatGPT Images or upload an existing image, then describe the change you want.”11
Many current tools mix both. Midjourney’s Edit Model, for versions 8.1 and 8.2, can “Generate new images using up to 4 reference images (replacing Omni Reference and Character Reference).”12 Black Forest Labs’ FLUX.2 editing lets you “Reference multiple images simultaneously”.13
Two close cousins: inpainting changes only an area you mark, and Hugging Face calls it “a more precise tool for making specific changes” than image-to-image.14 ControlNet borrows only a structure, such as a depth map or an outline; Hugging Face notes it “generates an image that preserves the information in it”.1
Want the same layout in a new look? Start from the picture. Want the same character in a new scene? Reference it. Want one thing changed? Edit by instruction.
How do you get a good result?
The plain point: change one thing at a time, say what must stay, and check the result closely.
- Name what to keep. OpenAI’s prompting guide says: “For edits, say ‘change only X’ and list the details to preserve, such as identity, geometry, layout, lighting, or labels.”15
- One change per turn. OpenAI’s ChatGPT guide advises: “Adjust one element at a time so the composition and other important details do not drift.”16
- Expect drift over many rounds. The same OpenAI guide warns: “Repeated edits can still change details you intended to preserve.”15 In this site’s use, going back to the original after a few rounds often beats editing an edit of an edit.
- Keep the words and the picture in agreement. In this site’s use, a prompt that contradicts the picture (blonde hair over a dark-haired photo) usually wins over the picture; how to use reference images covers why.
The second prompt carries no instructions about your sketch on purpose. Midjourney’s own advice for reference images is to “Focus on content, not instructions”: describe what you want to see.17 More edit wording is on image edit prompts that change only one thing.
What can go wrong?
The plain point: detail drift, plastic-looking results, and pictures you had no right to use.
- Small details change. Even a mask is not exact: OpenAI says its model uses a mask “as guidance, but may not follow its exact shape with complete precision.”18 Check faces, hands and any text.
- Waxy, over-smoothed results. Repeated passes can flatten texture; see why AI images look plastic.
- Real people. Putting someone’s photo into a scene they were never in can make a deepfake. OpenAI’s ChatGPT guide asks you to “confirm that you have permission to use their likeness.”16 Google’s Gemini help adds: “Editing images isn’t available to users under 18.”19
- Someone else’s artwork. Starting from another person’s image raises ownership questions; who owns what you generate explains what is known.
Questions people ask
Is image-to-image the same as img2img?
Yes. Img2img is the short name, taken from the img2img.py script in the original Stable Diffusion code (source above).
What strength should I use for image-to-image?
No vendor gives one right number; it depends on the model and how much change you want. Hugging Face’s software starts at 0.8 and ComfyUI’s KSampler at 1.0 (sources above). In this site’s use, try a low and a high value on the same picture and compare.
Why does the face change even at low strength?
Because the whole picture is noised and redrawn, so no pixel is copied straight through (Hugging Face and SDEdit, above). For a face that must stay the same, edit only the area around it, or use a face reference; see character consistency.
Can I turn a photo into a video the same way?
That is a different job, called image-to-video: the picture becomes the first frame of a moving clip. See image-to-video meaning and how to turn a photo into a video with AI.
Where to go next
- “Just the definition.” What is img2img?
- “I want the same face or style every time.” How to use reference images and character consistency.
- “I only want one thing changed.” Image edit prompts that change only one thing and inpainting.
- “I run models on my own computer.” ComfyUI explained and ControlNet.
- “What do the other words mean?” Diffusion model, CFG scale, style reference and the AI glossary.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does classic image-to-image do to your picture before redrawing it?
- It adds noise, then cleans it up towards your prompt
- It copies it pixel by pixel
- It turns it into a short video clip
Show the answer
Answer: It adds noise, then cleans it up towards your prompt. The page explains that your picture is noised and then denoised towards the prompt, which is why it keeps the layout but changes the look.
2. What happens when you set strength (or denoise) close to 1?
- The result stays almost identical to your picture
- Your starting picture barely matters
- The picture is only resized
Show the answer
Answer: Your starting picture barely matters. The page quotes Hugging Face: at a strength of 1.0 the initial image is more or less ignored.
3. Which way of working suits changing only the background and keeping everything else?
- Classic img2img at high strength
- A style reference
- An instruction edit that says what to change and what to keep
Show the answer
Answer: An instruction edit that says what to change and what to keep. The page says instruction edits change your picture in place, and vendor guides advise naming what to change and listing what to preserve.
- 1Hugging Face, Image-to-image, Diffusers documentation, undated, read at source 19 Oct 2026“Image-to-image is similar to text-to-image, but in addition to a prompt, you can also pass an initial image as a starting point for the diffusion process.” ↑a“The initial image is encoded to latent space and noise is added to it. Then the latent diffusion model takes a prompt and the noisy latent image, predicts the added noise, and removes the predicted noise from the initial latent image to get the new latent image. Lastly, a decoder decodes the new latent image back into an image.” ↑b“a higher strength value gives the model more ‘creativity’ to generate an image that’s different from the initial image; a strength value of 1.0 means the initial image is more or less ignored” and “a lower strength value means the generated image is more similar to the initial image” ↑c“if the num_inference_steps is 50 and strength is 0.8, then this means adding 40 (50 * 0.8) steps of noise to the initial image and then denoising for 40 steps to get the newly generated image.” ↑d“use a combination of low strength and low guidance_scale to generate an image that resembles the initial image but is not as strictly bound to the prompt.” ↑e“The conditioning image can be a canny image, depth map, image segmentation, and even scribbles! Whatever type of conditioning image you choose, the ControlNet generates an image that preserves the information in it.” ↑f
- 2Meng et al., SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations, arXiv:2108.01073, submitted 2 Aug 2021, as cited on this site’s img2img page, read at source 12 Oct 2026“SDEdit first adds noise to the input, then subsequently denoises the resulting image through the SDE prior to increase its realism.” ↑
- 3CompVis, stable-diffusion, GitHub README, undated, as cited on this site’s img2img page, read at source 12 Oct 2026run as scripts/img2img.py, “By using a diffusion-denoising mechanism as first proposed by SDEdit”. ↑
- 4Hugging Face, Image-to-image (StableDiffusionImg2ImgPipeline), Diffusers API documentation, undated, as cited on this site’s img2img page, read at source 12 Oct 2026“strength (float, optional, defaults to 0.8)” … “Must be between 0 and 1.” ↑
- 5Comfy Org, KSampler, ComfyUI documentation, undated, read at source 19 Oct 2026denoise, “Denoising strength - 1.0 for full denoising, lower values preserve original structure, suitable for image-to-image”; minimum 0.0, maximum 1.0, default 1.0. ↑
- 6Midjourney, Image Prompts, as cited on this site’s Midjourney cheat sheet, read at source 12 Oct 2026 in a browserchart for Version 8.1, Version 7 and Niji 7: “Image Weight Default 1 1 1”; “Image Weight Range 0 - 3 0 - 3 0 - 2”. ↑
- 7Midjourney, Style Ref v2, update post, 15 Mar 2024, as cited on this site’s style reference page, read at source 12 Oct 2026“similar to image prompts but focuses solely on the style.” ↑
- 8Ye et al., Tencent AI Lab, IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models, arXiv, Aug 2023, as cited on this site’s reference images page, read at source 17 Sep 2026“an effective and lightweight adapter to achieve image prompt capability for the pretrained text-to-image diffusion models.” ↑
- 9Brooks, Holynski and Efros, InstructPix2Pix: Learning to Follow Image Editing Instructions, arXiv:2211.09800, submitted 17 Nov 2022, read at source 19 Oct 2026“We propose a method for editing images from human instructions”. ↑
- 10Google AI for Developers, Nano Banana image generation, undated, as cited on this site’s image edit prompts page, read at source 12 Oct 2026“Provide an image and describe your change. The model will match the original image’s style, lighting, and perspective.” ↑
- 11OpenAI Help Center, Images in ChatGPT, as cited on this site’s character consistency page, read in a browser 12 Oct 2026“To edit an image, select an image generated by ChatGPT Images or upload an existing image, then describe the change you want.” ↑
- 12Midjourney, Edit Model, as cited on this site’s character consistency page, re-read directly at source 13 Oct 2026“Generate new images using up to 4 reference images (replacing Omni Reference and Character Reference).” and “The Edit model is compatible with Midjourney versions 8.1 and 8.2.” ↑
- 13Black Forest Labs, FLUX.2 Image Editing, as cited on this site’s character consistency page, read at source 13 Oct 2026“Reference multiple images simultaneously - up to 8 via API, up to 10 in the playground.” ↑
- 14Hugging Face, Inpainting, Diffusers documentation, undated, as cited on this site’s inpainting page, read at source 1 Oct 2026“You can think of inpainting as a more precise tool for making specific changes and image-to-image has a broader scope for making more sweeping changes.” ↑
- 15OpenAI, Image prompting, undated, as cited on this site’s image edit prompts page, read at source 12 Oct 2026“For edits, say ‘change only X’ and list the details to preserve, such as identity, geometry, layout, lighting, or labels.” ↑a“Repeated edits can still change details you intended to preserve.” ↑b
- 16OpenAI, ChatGPT Learn, Image generation, undated, as cited on this site’s image edit prompts page, read at source 12 Oct 2026“Adjust one element at a time so the composition and other important details do not drift.” ↑a“confirm that you have permission to use their likeness.” ↑b
- 17Midjourney, Style Reference, as cited on this site’s Midjourney cheat sheet, read at source 12 Oct 2026 in a browser“Focus on content, not instructions - Use your text prompt to describe what you want to see, not how Midjourney should modify the reference image.” ↑
- 18OpenAI, Image generation, undated, as cited on this site’s image edit prompts page, read at source 12 Oct 2026“The model uses the mask as guidance, but may not follow its exact shape with complete precision.” ↑
- 19Google, Gemini Apps Help, Generate & edit images with Gemini Apps, undated, as cited on this site’s image edit prompts page, read at source 12 Oct 2026“Editing images isn’t available to users under 18.” ↑
Read at source 19 Oct 2026: Hugging Face, Image-to-image (Diffusers documentation); Comfy Org, KSampler; Brooks, Holynski and Efros, InstructPix2Pix (arXiv, 2022).
Reused with the read dates recorded on this site’s img2img, inpainting, style reference, reference images, character consistency and image edit prompts pages and the Midjourney cheat sheet: SDEdit (Meng et al., arXiv, 2021); CompVis stable-diffusion; Hugging Face Diffusers API and Inpainting docs; IP-Adapter (Ye et al., arXiv, 2023); Midjourney docs and update posts; Google’s Nano Banana and Gemini Apps help pages; OpenAI’s image prompting, image generation, ChatGPT Learn and Help Center pages; Black Forest Labs’ FLUX.2 editing docs.
The chalk-drawing comparison, the example prompts and the advice marked “this site’s use” or “this site’s reading” are this site’s own, not claims from these sources. Nothing on this page was tested as a measured result.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK