If Perchance image to image is not working, the problem is usually the source image rather than the prompt. Oversized or unsupported files are rejected on upload, a low strength setting produces output that looks unchanged, and a source image that trips the input filter returns nothing. Identify which of the three you have before changing anything.
Image to image fails differently from text to image
This is the point that most troubleshooting advice misses, and it is the reason people waste hours. Text to image has one input: your words. Image to image has three: your words, your source image, and the strength setting that decides how much of the source survives. Two of those three are things text to image never has to deal with, and both introduce failure modes that no amount of prompt editing will fix.
So before you rewrite your prompt for the fifth time, work out which stage is actually breaking. There are only three, and they look completely different from each other.
The three stages, and how each one fails
Stage one is the upload. Your file has to be accepted, which means the right format and a manageable size. Failure here is usually visible: the image never appears, or it appears and then vanishes, or the interface simply does not acknowledge it.
Stage two is the input check. An accepted image still has to pass whatever screening the generator applies to uploaded content. Failure here is quiet. Everything looks fine and then nothing comes back, which is why it gets misdiagnosed as an outage more than any other cause.
Stage three is the transformation itself. The image is accepted, it passes, generation runs, and the result looks essentially identical to what you put in. That is not a failure of the tool. That is the strength setting doing exactly what you asked, which we cover in depth in our image to image guide.
Symptom to cause, at a glance
| What happens | Stage | Most likely cause | First action |
|---|---|---|---|
| Image never appears after selecting it | Upload | File too large or wrong format | Convert to JPG or PNG, reduce dimensions |
| Upload stalls partway | Upload | Slow or unstable connection | Retry on a stronger connection |
| Upload works, generation returns nothing | Input check | Source image screened out | Test with a plainly harmless image |
| Output looks identical to the source | Transform | Strength set too low | Raise strength substantially |
| Output ignores the source entirely | Transform | Strength set too high | Lower strength |
| Text to image works, image to image never does | Upload | Extension blocking the upload path | Test in a private window |
Upload problems come first
File size is the usual culprit
Modern phone cameras produce enormous files. A single photo can be several times larger than any browser based generator wants to handle, and the failure is often silent: you pick the file, something churns, and nothing happens. Before assuming the tool is broken, shrink the image. Resizing the longest edge down to something in the region of a thousand to fifteen hundred pixels solves an astonishing proportion of image to image complaints on its own, and costs you nothing in output quality because the generator is not going to use that extra resolution anyway.
Format matters more than people expect
Standard JPG and PNG are safe. Newer or less common formats are not universally handled, and your phone may be producing one of them without telling you. Screenshots, live photos and images saved from messaging apps are common offenders. Convert to a plain JPG and retry before investigating anything else. This is a thirty second test that eliminates an entire category of cause.
Very large dimensions cause their own trouble
Separate from file size, extreme pixel dimensions can fail even when the file is small. Panoramas, long screenshots and stitched images are the usual examples. A heavily compressed image can be small on disk and still be far too wide for the generator to handle, so file size alone is not a reliable guide. Crop to something closer to a normal aspect ratio and retry.
Transparency and unusual colour data
PNG files with transparent regions sometimes behave unpredictably, because the generator has to decide what those transparent pixels mean before it can work with them. If your source has a transparent background, flatten it onto a plain white or black background first and upload that instead. The same caution applies to images saved with unusual colour profiles, which are common in files exported from design software. Opening the file and re-saving it as an ordinary photograph style JPG normalises both problems at once, and takes less time than working out which of the two you actually had.
When the upload path itself is blocked
If text to image works perfectly and image to image never gets off the ground, suspect your browser rather than the file. Some privacy extensions interfere with file handling specifically. Open the generator in a private window, where extensions are usually disabled, and try the same upload. If it works there, you have found your cause and can re-enable extensions one at a time to identify which one.
The source image itself can be rejected
This is the cause people find hardest to accept, because it produces no explanation at all. Generators screen uploaded images as well as prompts. An image that trips that screening produces nothing, and often produces nothing silently. You are left staring at an interface that appears to be working and concluding the service must be down.
The diagnostic is straightforward. Upload something unambiguously mundane, a landscape photo or a picture of a chair, keep everything else the same, and generate. If the mundane image produces output and your original does not, the source image is the variable, not the tool, not your connection and not your prompt. That is a clean result and it saves you from chasing a fault that does not exist.
Filters on free tools are also conservative by design, and they are tuned to over reject rather than under reject. That is a defensible engineering choice and an irritating user experience. If you are consistently hitting it with source images you consider perfectly reasonable, you have reached the practical ceiling of a free browser tool and a service built for this work, such as AI Nudez, will handle input images more predictably. Our comparison of free image to image generators lays out what the free tier options can and cannot realistically do.
When nothing visibly changes
Strength is the setting that matters
Image to image works by partially destroying your source image with noise and then reconstructing it under the guidance of your prompt. The strength setting, sometimes labelled denoise, controls how much gets destroyed. Set it low and very little is destroyed, so very little changes, and the output is your source image with a slightly different texture. That is not a bug. That is the setting doing precisely what it was told.
If your output looks unchanged, raise strength substantially rather than nudging it. Small increases produce small differences, which is exactly the situation you are trying to escape. Make a large jump, see what happens, then work back down toward the balance you want. Two or three attempts with big steps beats ten attempts with small ones.
The opposite failure looks like a different image entirely
Push strength too far and the source stops mattering. Enough of it is destroyed that the model is effectively working from your prompt alone, and you get an unrelated image. People read this as the upload having failed, when in fact the upload worked perfectly and the setting overrode it. If your output ignores the source, the fix is to go down, not up.
Your prompt still has to describe the destination
A common mistake is describing what the source image already contains. The prompt in image to image should describe what you want to end up with, including the things you are changing. If your prompt only restates what is already there, the model has no instruction to work from and will return something very close to the source at any moderate strength.
When you only want to change one region
If you are trying to alter one part of an image while leaving the rest alone, image to image is the wrong tool and will keep disappointing you, because it applies change across the whole frame. What you want is inpainting, which restricts change to a region you define. Our inpainting workflow guide covers that properly. Similarly, if you need to preserve a specific pose or composition while changing everything else, that is a job for ControlNet rather than raw strength adjustment.
Quality problems that are not faults
Some results look broken but are working as designed. Output that is softer or blurrier than the source usually means the source was higher resolution than the generator works at, so detail was lost on the way down and not restored on the way back up. That is expected, and our blurry output guide covers the ways around it. Faces and hands degrading during transformation is also normal at moderate strength, because those are the areas the model reconstructs least reliably.
Testing methodically instead of guessing
Image to image has more variables than most people hold in their head at once, which is why the usual approach, changing several things and trying again, produces so little progress. A short, deliberate method beats it comfortably.
Establish a known good baseline first
Before you troubleshoot your actual project, get one successful transformation of any kind. Use a plain photograph, a plain prompt, and a strength value somewhere near the middle of whatever range the interface offers. If that produces a changed image, you have proven that uploads work, screening passes, and transformation runs. Every subsequent failure can then be attributed to whatever you changed away from that baseline, which is enormously more informative than starting from a broken state and guessing.
Change exactly one thing per attempt
From the baseline, swap in your real source image but keep the plain prompt and the same strength. If that fails, the source image is your problem. If it succeeds, swap in your real prompt next. If that fails, the prompt is your problem. Then adjust strength. Three attempts, arranged this way, will tell you more than twenty attempts arranged randomly, and the whole sequence takes only a few minutes.
Write down what you tried
This sounds fussy and it is not. Strength values in particular are easy to lose track of after several attempts, and people routinely end up re-testing a combination they already ruled out twenty minutes earlier. A few lines in a notes file prevents that entirely, and it also gives you a record of the settings that eventually worked, which is worth far more than the single image you were trying to make.
The fix order
Work through this and stop when it works. Change one thing at a time, because changing three and succeeding tells you nothing.
Start by resizing your source image down and converting it to plain JPG. Then test the upload with a mundane image to separate an upload fault from an input screening one. Then, if generation runs but output looks unchanged, make a large increase to strength. Then rewrite the prompt so it describes the destination rather than the source. Then, if image to image never works at all while text to image does, test in a private window to rule out extensions. If every one of those fails on multiple browsers and devices, the fault is not local, and the broader Perchance troubleshooting overview covers what else could be going on.
Knowing when the free tool is the limit
Image to image asks more of a free browser tool than text to image does. Uploads consume bandwidth, source images need screening, and the whole pipeline has more places to break. That is why the free experience here is noticeably rougher than plain text prompting, which is worth knowing before you conclude that you are doing something wrong. Our writeup on what Perchance actually offers is honest about where those limits sit.
If image to image is central to how you work, rather than an occasional experiment, a purpose built service will save you more time than any amount of troubleshooting. AI Nudez handles uploads on its own infrastructure and gives clearer feedback when something is rejected, which is the single most valuable difference in practice. Our list of Perchance alternatives covers the rest of the field if you want to compare first.
Frequently asked questions
Why is Perchance image to image not working when text to image is fine?
Because image to image adds two inputs that text to image does not have: an uploaded file and a strength setting. The most common causes are a source file that is too large or in an unsupported format, and an extension interfering with file uploads specifically. Resize and convert your image to plain JPG, then test in a private window.
What image size should I upload to Perchance?
Resize the longest edge down to roughly a thousand to fifteen hundred pixels before uploading. Straight from a phone camera, files are often far larger than a browser based generator handles comfortably, and oversized uploads frequently fail silently. You lose nothing in output quality, since the generator works at a lower resolution than your camera produces anyway.
Why does my output look exactly like the image I uploaded?
Your strength or denoise setting is too low. That setting controls how much of the source is destroyed before reconstruction, and a low value means very little changes. Raise it substantially rather than nudging it up slightly, then work back down toward the balance you want. Two large steps beat ten small ones.
Why does my output ignore the source image completely?
The opposite problem: strength is set too high. Enough of the source has been destroyed that the model is effectively generating from your prompt alone. Lower the setting. Many people mistake this for the upload having failed, when in fact the upload worked correctly and the strength value overrode the source entirely.
Can Perchance reject the image I upload?
Yes, and it can do so without any visible explanation, which is why this cause is so often misread as a service outage. To test it, upload something unmistakably mundane like a landscape photo, keep everything else the same, and generate. If the mundane image produces output and yours does not, the source image is the variable.
Which file formats work best for image to image?
Plain JPG and PNG are the safest choices. Newer or less common formats are not handled universally, and phones increasingly save in them without making it obvious. Screenshots and images pulled from messaging apps are frequent offenders. Converting to a standard JPG before uploading is a thirty second test that removes an entire category of failure.
How do I change only part of an image?
Not with image to image, which applies change across the whole frame regardless of what you want. The tool for region specific edits is inpainting, where you define the area to change and the rest is preserved. Trying to force selective edits through strength adjustment produces the frustrating result of everything changing slightly and nothing changing enough.
Why is my transformed image blurrier than the original?
Because your source was likely higher resolution than the generator works at internally. It was scaled down for processing and the lost detail was not restored on the way back out. This is expected behaviour rather than a fault. Upscaling afterwards is the normal remedy, and starting from a source closer to the working resolution helps too.
