How to Fix Persistent Shadow Loops & Asymmetrical Skin Tones on Flow (Flow - NB2 & Pro)

Hello everyone,

I’m Vel and today I’ll share one of my latest experiences. I’d like to hear more about your ideas, and perhaps my solution can help some people who are dealing with this issue.

For quite some time, I had been getting distorted results on my characters — asymmetrical skin colors, unnatural lighting, and excess shadows that seemed to loop no matter what I tried. I’m working with this method on Flow, using Nano Banana 2 and Pro: reference image(s) + text → new images.

To solve these issues, I tried more than a hundred prompts, including my own prompts and prompts generated by ChatGPT, Gemini, and Claude. The common thread among all the prompts was that they all failed, regardless of the model used — the results showed almost no change. At one point, I even tried manually cutting out the parts I wanted corrected, thinking that would prevent the model from replicating the original structure, but that didn’t work either.

However, after that method also failed, I found two workarounds that actually fixed it:

Two workarounds that fixed it for me:

  1. Start a new, blank project and upload only the corrupted images there.
  2. In the existing project, add these two lines at the start of your prompt: “IGNORE ALL PROJECT MEMORY” and “FORGET ALL PROJECT MEMORY”

After either workaround, the model will begin correctly altering skin color, lighting, and shadows — including breaking the shadow loop that wouldn’t budge before. From there, refine your prompt to match the exact result you’re looking for.

Hope this helps anyone stuck in a similar prompt loop! Feel free to ask any questions or share if you’ve found other cool tricks.

Do you have any examples of the “shadow loops”? I have been seeing something similar in skin and clothing when the shot is of a model centered in a studio style background. I don’t have a generic example I can share right now.

Curious what you’re seeing and if it’s the same issue.

(Before)
Sure, here’s a before / after comparison. The dark shadow on the left side of her face in the top image is a perfect example of what I mean. Without applying the memory bypass and context workarounds I mentioned, the models continuously lock onto those pixels and regenerate them across iterations—regardless of how many different prompts I tried.


(After)
(Had to use another message because I can’t upload 2 images to one message.)