Editing Images with Gemini 3.1 Flash Image

Nicknamed “Nano Banana 2,” Google’s new flagship GenAI model is fantastic for editing images without changing existing details outside the scope of the edit (unlike some other premier models—cough, Flux.2 Max, cough).

Okay, without changing them much. More on that later.

Need to do a simple edit to your character concept art when the weather turns warm?

NB2 can also change the material an element in an image is made of and apply patterns to a surface—all based on reference images.

You can transfer materials to pretty much literally anything. Like, I needed an access panel for Gus, so…

What impressed me more is NB2 can apply a pattern to draped fabric. Updating Wulan’s original AI-slop scarf was a (relative) breeze.

To start, I sourced an authentic Batik-patterned design.

Then with NB2 (and no little TLC in Photoshop), I produced the old, worn, stained, and sun-bleached gold and indigo version from the novel.

Finally, it was a one-step operation in NB2 to swap the scarf with the culturally-genuine article using this natural language prompt:

Create a new image: the scarf exactly as shown in @img2 — same drape, folds, wrapping, shape, size, and frayed edges — but made of the fabric shown in @img1. Render @img1's batik pattern, faded indigo ground, antique gold linework, fading, and staining and dirt and oil stains as depicted, onto the scarf, following the scarf's folds so it reads as real draped cloth. Use the woman in the jacket from @img2 on a solid white background for the new image, preserving the framing and angle of her body, and the lighting from @img2 unchanged. Photorealistic, 8K UHD, fabric taking the scene's light across every fold.

Not only can NB2 do material and pattern transfer—it can transfer text.

Take my Doors to the Stars protagonist’s name in trade pidgin for example:…

And say you wanted to have her tag her faithful recon drone Gus with her name on one of his repair panels. But not just a lame Photoshop blend mode. Like she really wrote in on the panel with paint that has thickness and texture.

Impossible, you say?

Au contraire, mon ami!

Only problem is it can’t quite figure out how to curve a line of alien trade pidgin to fit a shape (and believe me, I tried).

But 30 seconds in Photoshop with warp transform solved that problem easy enough.

I go into more detail (with example prompts) in my follow-up tutorial.

Reference Images

The topic of reference images deserves a completely separate tutorial, but here’s just a glimpse of what you can do with them:

I’m sorry, but that’s just damned cool (I added the locking lugs intentionally in a subsequent iteration).

Far more advanced compositing is possible, and NB2 supports up to fourteen reference images for truly complex operations (I recommend using JSON instead of natural language for advanced prompts). I’m still experimenting with creating composite images from individual parts. Stay tuned.

Google Nano Banana 2’s Undocumented Power Tool: Annotated Reference Images

But the most interesting thing I’ve discovered is that NB2 can understand annotations in reference images. I’m still testing and validating the extent of the capabilities, but the following test is reproducible across multiple generations, so it wasn’t luck.

This works for defining sizes and shapes too. Say you want to add a panel seam in a metal hull. Getting GenAI to nail the exact size you want is impossible with a prompt. But with an annotated reference image…

The following example was one of the more complex annotated references I’ve attempted. Red circle for the size of the transparasteel dome (because the model was getting really confused about the metal ring being a boundary) and yellow circle for the impact point and general size of the divot.

I tested both a JSON and natural language prompt, and honestly I can’t say I necessarily got more consistently accurate and reproducible results from JSON. I didn’t bother logging the hits and misses over all the generations I tested, so I can’t give hard numbers, but it was essentially a toss-up with maybe somewhat better averages from JSON. 

What did land solidly in the “hits” category was the annotated reference image. 90% of the time the damage was contained in the red circle and the impact point was right where the yellow circle is—and the divot was generally about that size. Also worth noting NB2 seems to understand filled shaped better than outlines—but I haven’t proved that definitively yet.  

{
 "task": "Edit @img2. Add blaster bolt damage to the existing transparasteel dome surface only. Single impact point. Do not re-render the dome. Do not alter the dome surface in any way except to add the cracks. Minimal change only.",
 "hard_constraint": "Do not alter the dome surface in any way except to add the crack damage described below. Do not re-render the dome. Do not add, change, or enhance reflections on the dome. Do not change the dome's transparency, color, tone, or appearance. The dome looks exactly as it does in @img2. Only the cracks are new.",
 "base_image": "@img2",
 "reference_images": [
   {
     "tag": "@img1",
     "role": "annotation reference only — do not render anything from this image",
     "annotation_one": "filled red circle marks the exact size and boundary of the transparasteel dome inside the lens housing ring",
     "annotation_two": "filled yellow circle marks the exact location of the impact and size of the divot in the transparasteel dome"
   }
 ],
 "change": {
   "target": "transparasteel dome over the central lens — the area marked by the filled red circle in @img1",
   "hard_boundary": "ALL damage — every crack, every fracture line, the divot — is fully contained within the boundary marked by the filled red circle in @img1. Nothing crosses outside this boundary. Zero exceptions.",
   "material_properties": "Transparasteel is not glass. It does not shatter or produce powdery white fracture edges. Cracks are clean, sharp-edged, and precise — like stressed polycarbonate or cracked carbon fiber. Tough, reluctant damage, not explosive. It has a matte anti-reflective finish, without the reflectivity of glass.",
   "damage": {
     "impact_count": "ONE impact point only at the location defined by the filled yellow circle in @img1. There is exactly one divot and one set of cracks. Do not generate two impact points.",
     "single_impact": "one small divot or dimple the size of the filled yellow circle in @img1 where the transparasteel deformed inward under impact — the surface is intact and unpenetrated — no hole, no penetration, no puncture. Material compressed but did not fail.",
     "radial_cracks": "several short clean sharp-edged cracks radiating outward from the single impact point, contained within the dome boundary",
     "main_crack": "one large crack running diagonally from the location indicated by the filled yellow circle in @img1 to the extreme upper camera-left edge of the dome boundary denoted by the filled red circle in @img1",
     "crack_edges": "subtle stress iridescence where the material's internal structure is exposed — faint, not dramatic",
     "delamination": "slight depth along the deeper cracks — layers separating minimally",
 },
 "preserve": {
   "dome_surface": "exactly as it appears in @img2 — reflectivity, color, tone, transparency, size, and appearance unchanged. Do not add or enhance any reflections.",
   "lens": "unchanged",
   "hull": "unchanged",
   "four_optical_sensors": "unchanged",
   "angle_and_framing": "unchanged",
   "background": "white, unchanged"
 },
 "rendering": {
   "quality": "8K UHD photorealistic — photographic, not CGI",
   "white_balance": "neutral white light",
   "color": "flat photographic, no color grading, no cinematic styling",
   "camera": "Phase One IQ4 150MP, 35mm f/8, deep depth of field"
 }
}

Edit @img2. Add blaster bolt damage to the existing transparasteel dome surface only. Single impact point. Do not re-render the dome. Do not alter the dome surface in any way except to add the cracks. Minimal change only.

@img1 is an annotation reference only — do not render anything from it. The filled red circle in @img1 marks the exact size and boundary of the transparasteel dome inside the lens housing ring. The filled yellow circle in @img1 marks the exact location of the impact and the size of the divot.

Do not alter the dome surface in any way except to add the crack damage described below. Do not re-render the dome. Do not add, change, or enhance reflections on the dome. Do not change the dome's transparency, color, tone, or appearance. The dome looks exactly as it does in @img2. Only the cracks are new.

All damage is contained within the boundary marked by the filled red circle in @img1 — every crack, every fracture line, the divot. Nothing crosses outside this boundary. Zero exceptions.

Transparasteel is not glass. It does not shatter or produce powdery white fracture edges. Cracks are clean, sharp-edged, and precise — like stressed polycarbonate or cracked carbon fiber. Tough, reluctant damage, not explosive. It has a matte anti-reflective finish, without the reflectivity of glass.

There is ONE impact point only, at the location defined by the filled yellow circle in @img1. Do not generate two impact points. At that location: one small divot the size of the filled yellow circle where the transparasteel deformed inward under impact — the surface is intact and unpenetrated, no hole, no penetration, no puncture. Material compressed but did not fail. Several short clean sharp-edged cracks radiate outward from the single impact point, contained within the dome boundary. One large crack runs diagonally from the impact point to the extreme upper camera-left edge of the dome boundary marked by the filled red circle in @img1. Along the crack edges, subtle stress iridescence where the material's internal structure is exposed — faint, not dramatic. Along the deeper cracks, slight delamination — layers separating minimally.

Preserve everything else exactly as it appears in @img2: the dome surface reflectivity, color, tone, transparency, size, and appearance unchanged — do not add or enhance any reflections. The lens, hull, four optical sensors, angle, framing, and white background are all unchanged.

8K UHD photorealistic quality — photographic, not CGI. Neutral white light white balance. Flat photographic color, no color grading, no cinematic styling. Shot on Phase One IQ4 150MP, 35mm f/8, deep depth of field.

Google’s own documentation doesn’t mention this capability, and as far as I know I’m the only one who’s reported discovering it.

Update: Took some searching, but aifire.com documented this back in November, 2025. I just happened to discover it independently half a year later.

Far more experimentation is required, but initial tests have even proven successful with text labels and arrows itemizing specific parts.

Which is seriously cool.

The Gotchas

It’s worth mentioning NB2 can’t really handle more than one or two specific edits in the same operation, so multiple iterations are required for complex multi-element edits, but it downsamples reference images aggressively and doesn’t maintain pixel-perfect fidelity when “preserving” the unedited potions of images, so multiple edit iterations will degrade your image quality significantly. Your UHD photo-realism will blur into CGI mush, your colors will shift into bland pottage, and your subject will inexplicably change body proportions. My workaround is old school Photoshop kit bashing—I’ll make edits with NB2 and then mask the updates and apply them to a high fidelity master image in PS just like we did way back in 2021. If I find a better workaround I’ll let you know.

Regardless, the editing features are ultimately a major win.

Also worth mentioning it introduces a weird violet cast to images, so you’ll have to be on guard and do frequent color correction because if you iterate on an iteration of an iteration, etc., it’ll eventually turn purple.

Also, also, it can’t resist pushing all human subjects to whiteness (like all GenAI), so if you’re working with an image of a non-Caucasian (I’ve been told they exist), you’ll have to be on guard for colonization too.

I’ve posted documentation on crafting prompts for image editing in NB2. It’s a living document, and I’ll keep editing it as I learn more, so be sure to check back.