Img2img in one picture: the original is partially dissolved, then rebuilt. How far you let it dissolve is the whole game.
Img2img in one picture: the original is partially dissolved, then rebuilt. How far you let it dissolve is the whole game.
Text-to-image starts from pure noise and a prayer. Img2img starts from a picture you already have, and one slider decides whether the model gives it a gentle polish, repaints it in a new style, or tears it down to the foundation and builds something new on the same footprint.
Hey friends. Today we are talking about the tab most people click once, get a weird result from, and never open again: img2img. Which is a shame, because image-to-image generation is quietly one of the most powerful workflows in all of AI art. It is how you rescue an image that is eighty percent perfect instead of rerolling and losing it. It is how you turn a clumsy phone sketch into a finished render. It is how you repaint a photo as a watercolor, an oil painting, or an anime frame while keeping the composition you loved. And it is how you generate ten disciplined variations of a winner instead of ten random strangers.
The reason most people bounce off img2img is that it has one setting that changes everything, and nobody explains it. That setting is denoising strength. So that is where we will start: what img2img actually does under the hood, how the strength slider trades faithfulness for freedom, the three working sweet spots I reach for constantly, and then the fun stuff, style transfer, sketch-to-render, and the variation loop that turns near-misses into keepers.
Here is the mental model that makes everything else click. A diffusion model generates pictures by starting from pure random noise and removing that noise step by step until an image emerges, steered the whole way by your prompt. Text-to-image runs that process from one hundred percent noise, which is why every generation is a fresh gamble.
Img2img changes only the starting point. Instead of beginning from pure noise, the model takes your input image, adds a controlled amount of noise to it, and then denoises from there. If it only added a little noise, most of your original survives and the model can only nudge details. If it added a lot, your original is nearly erased and only its ghost, the rough composition, the placement of light and dark, the general shapes, remains to guide the rebuild. Your prompt still steers everything, but now it is steering a renovation instead of a construction from empty ground.
That is the entire trick. Every img2img workflow, from subtle cleanup to full style transfer, is just choosing how much of the original to destroy before rebuilding.
Denoising strength runs from 0 to 1, and it answers one question: how much freedom does the model get? At 0, nothing happens and you get your input back. At 1, the input is effectively ignored and you are back to text-to-image. Everything interesting lives in between, and the behavior is not linear, it has zones.
| Strength | What survives | What it is for |
|---|---|---|
| 0.1 to 0.3 | Almost everything: composition, pose, face, colors | Cleanup and polish. Smoothing artifacts, adding fine texture, gently sharpening a soft render. |
| 0.4 to 0.6 | Composition, pose, and palette; surfaces get repainted | The restyle zone. Changing the medium, the outfit details, the rendering style, while the picture stays recognizably itself. |
| 0.7 to 0.85 | Rough layout and the big shapes only | The reimagine zone. Using the original as a loose compositional suggestion for a substantially new image. |
| 0.9 and up | Almost nothing | Barely different from a fresh generation. Rarely what you want in img2img. |
The habit to build: decide which zone your goal lives in before you touch anything else. "I love this image, fix the mushy hands" is a 0.2 to 0.3 job, and honestly often a job for masked inpainting instead. "Same picture but as an oil painting" is 0.4 to 0.6. "Keep this general layout but make it a completely different scene" is 0.7 plus. Most img2img frustration is someone doing a 0.5 job at 0.8, or wondering why nothing changed at 0.2.
The most satisfying img2img workflow is the restyle. Take an image whose composition you love, a photo, a render, one of your own generations, and repaint it in a different medium. The recipe is simple: load the image, set denoising strength in the 0.4 to 0.6 zone, and write your prompt as if you were describing the finished restyled image, leading with the new medium. "Watercolor illustration of a woman in a sunlit cafe, loose brushwork, soft paper texture" over a photo gives you that photo's exact framing reborn as a watercolor.
Three tips make restyles land. First, describe the destination, not the change. The model does not know what the input used to be, so "convert this to anime" works far worse than simply describing an anime frame. Second, keep the structural words from the original in your prompt, the pose, the setting, the lighting, so the model does not fight the image it was handed. Third, if the style is barely taking hold, raise strength in small steps of 0.05 rather than leaping, because the difference between 0.5 and 0.65 is often the difference between "same face, new brushwork" and "who is this."
Here is the workflow that feels most like magic. You do not need to draw well to compose well. Block out your idea as a crude sketch, and I mean crude, colored blobs for figure and background, a scribble for hair, a smear of orange where the sunset goes. Feed that into img2img at roughly 0.6 to 0.75 strength with a full prompt describing the finished image, and the model treats your blobs as a compositional blueprint and builds a real image on top of them.
This flips the usual power dynamic. Instead of begging the prompt lottery for the composition you want, you place every element yourself with five minutes of clumsy painting, then let the model do the rendering. Color your sketch deliberately, because the model reads those colors as palette instructions too. If the result respects your layout but looks too rough, run the output back through img2img at 0.3 to polish it. If the model steamrolled your layout, drop the strength a notch and try again. Pair this with the framing ideas from our camera and lens language guide and you are composing like a director instead of gambling like a slot player.
The most practical everyday use of img2img is the humble variation loop. You generated something eighty percent great. The instinct is to reroll, but a fresh roll throws away the eighty percent along with the twenty. Instead, send the winner to img2img, keep the same prompt, set strength around 0.3 to 0.45, and generate a batch. Every output is a sibling of your image, same composition, same character, same mood, with the details reshuffled. Somewhere in that batch is usually the version where the hands behave and the expression lands.
You can also steer while you vary. Tweak the prompt as you loop, warmer light, softer smile, and the change arrives without losing the picture. This loop stacks beautifully with seed control, which we covered in the seeds and reproducibility guide, and it is the backbone of a professional finishing pass before the final polish in our upscaling and color grading workflow.
A quick honesty section, because knowing the limits saves hours. If the problem is one small region, a hand, an eye, a stray object, masked inpainting beats whole-image img2img every time, since img2img at any meaningful strength will drift details everywhere, not just where you wanted. If you need the exact same face across many images, img2img alone will slowly mutate identity with every pass, so lean on the techniques in our character consistency guide. And if you find yourself at 0.9 strength trying to force a totally different image, stop, you are fighting the tool, and a fresh text-to-image prompt will serve you better. Tool choice is a skill of its own, and our complete guide to AI image generators covers which platforms expose real img2img controls and which hide them.
The img2img cheat sheet: 0.2 to polish, 0.5 to restyle, 0.7 to reimagine. Describe the destination image, never the edit. Move the slider in steps of 0.05 when you are close. And when only one small thing is broken, inpaint it instead of rolling the whole picture through the machine.
Img2img is the difference between generating images and directing them. Text-to-image will always be the fun slot machine, but the artists whose work looks intentional are the ones who treat a generation as a draft: sketch the composition themselves, restyle it until the medium sings, run disciplined variation loops on the near-misses, and polish the winner. All of that runs through one unassuming slider, and now you know what it actually does. Start at 0.5, watch what survives and what dissolves, and you will have an intuition for denoising strength within an evening.
If you want to keep building, lock your compositions with the seeds and controlled variation guide, repair the small stuff with the inpainting guide, and take your best loop survivor through the finishing workflow. You can see finished results across our galleries any time.
Happy generating, and send me your best before-and-after transformation!