edit-images

On the rail as nano-banana, nano-banana-2.1, nano-banana-lite, nano-banana-pro, gpt-image-2, gpt-image-2.5, flux-2-pro, flux-2-max, gpt-image-1.5, qwen-3.

When to use it

Changing what a picture shows by describing it: a new expression, outfit or sky colour, the light or the background, an object added or removed, or a logo or garment carried from one of up to six pictures onto another.

When to reach for something else

It holds at most six images, so a composite built from more sources takes more than one pass. There is no way to mark an exact region to change, so a boundary-locked edit needs a tool built around painting that area.

How to use it

1. Drop one or more pictures onto the card’s Images slot.
2. Write what you want in Prompt.
3. Pick a model in the pill at the top of the card, or keep Nano Banana Lite.
4. Press Run.

On the card

Control	What it does	Choices
Images	The pictures the tool works from. Drop them from the board or your computer, then refer to each one by number in the prompt.	Several pictures
Prompt	What you want, in your own words. The card sends it with your inputs.	Your words
Run	Starts the run. The number on it is what this run costs in credits, worked out before you press.
×1	Runs the card several times in a row with the same settings.	1 to 99

Models

The pill at the top of the card picks the model. Nano Banana Lite is the one a new card starts on.

Model	Good for
Nano Banana	Google’s Gemini 2.5 Flash Image for quick everyday edits: a plain instruction over up to six inputs for composition, and the engine 3D Prepare uses to lift a subject onto clean white, centred and re-lit for a 3D generator.
Nano Banana 2.1	The upgraded Nano Banana edit with a resolution ladder from 1K to 4K, a seed, and up to six inputs for composition; on the Blend in Scene pill it is a position-faithful pick that keeps a placed subject where you put it.
Nano Banana Lite	Google’s fast, cost-effective edit tier and the model the image-edit tool opens on: a plain instruction over up to six inputs, back at a fixed 1K.
Nano Banana Pro	Google’s Nano Banana Pro reached through Krea, the premium pick on the image-edit pill: multi-image edits over up to six inputs, referenced as image 1 and image 2 in the prompt, at 1K, 2K or 4K.
GPT Image 2	OpenAI’s GPT Image 2 on the image-edit pill: up to six @-tagged references plus a prompt, an optional mask, three quality tiers and output up to 4K, opened at medium quality and 1K.
GPT 2.5 Flare	Prompt-scoped edits on one to six reference images with an optional mask, holding subject, composition and background across successive rounds; the everyday variant, quicker than Sunburst at the same price.
GPT 2.5 Sunburst	The finer, slower variant of the same editor: one to six references, optional mask, edits scoped tightly to the instruction, for intricate detail that has to survive many rounds of revision at the same price as Flare.
Seedream 5 Pro	Instruction-led edits across up to ten reference images, and the photographic choice where it sits beside a stylising model: compositing a subject into a scene, recomposing one frame per platform, fusing objects.
Flux 2 Max	Black Forest Labs’ highest-fidelity FLUX.2 tier for generation and reference-guided editing: the strongest prompt adherence and most consistent multi-reference editing in the FLUX.2 line, up to six inputs tagged @ in the prompt.
Flux 2 Pro	The FLUX.2 workhorse across three tools: photoreal text-to-image with wide format and aspect control, reference-guided edits with up to six @-tagged inputs, and the engine Recolor uses to repaint a picture to an exact palette.
GPT Image 1.5	OpenAI’s GPT Image 1.5, the side-by-side alternative to GPT Image 2 on both the text-to-image and image-edit pills: one model for prompt-only generation and multi-image editing, quality tiered low to high.
P Image Edit	The quick, cheap pass in the Image Edit tool: an edit described in words across up to five pictures referenced as “image 1”, “image 2” and so on, back in about a second.
Uni-1.1	Reference-heavy edits where each extra picture has a named job (this palette, this lighting, this product) and the parts of the source you do not mention stay put; the standard tier, with optional web search.
Uni-1.1 Max	The higher-quality tier of the same reference-steered editor: identical controls (source, prompt, up to eight named references, web search) with better output, for the final pass when the standard tier is close but not clean enough.
Qwen Image 3	Edits that keep a face or identity intact while swapping an object, background or style, and edits that write or rewrite dense in-image text; one to three images, Chinese or English prompts.
FLUX 3	Several targeted edits to one picture in one run, each in its own box: replace one thing, remove another, put a reference picture into an area, and everything else stays as it was. Also edits from words with up to nine references.

Nano Banana

Not the one for: No seed, so a run cannot be reproduced exactly, and no resolution control, only aspect ratio. Not for region-locked edits or 4K output; its siblings on the pill carry the resolution ladder.

Behind the settings button:

Setting	What it does	Choices
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Match Input, Square 1:1, Vertical 2:3, Horizontal 3:2, Vertical 3:4, Horizontal 4:3, Vertical 4:5, Horizontal 5:4, Vertical 9:16, Horizontal 16:9, Horizontal 21:9
Output Format	The file type: PNG keeps the best quality and supports transparency, JPG is a smaller file, WebP is small and modern.	JPG, PNG

Nano Banana 2.1

Not the one for: Its default output is 2K, which costs more than 1K, so drop the resolution for a quick draft. Not a mask tool: a change confined to one painted region belongs in inpainting.

Behind the settings button:

Setting	What it does	Choices
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Auto, Square 1:1, Vertical 2:3, Horizontal 3:2, Vertical 3:4, Horizontal 4:3, Vertical 4:5, Horizontal 5:4, Vertical 9:16, Horizontal 16:9, Horizontal 21:9
Resolution	How big and detailed the result is. Higher looks sharper but costs more and takes longer.	1K, 2K, 4K
Seed	The starting dice roll. The same seed and settings give the same result every time; change it for a fresh variation, or leave it at -1 for a new random one.	-1 to 999999

Nano Banana Lite

Not the one for: Output is fixed at 1K with no resolution control, only aspect ratio, so anything that has to ship above 1K needs a sibling on the pill. No mask input.

Behind the settings button:

Setting	What it does	Choices
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Auto, Square 1:1, Vertical 2:3, Horizontal 3:2, Vertical 3:4, Horizontal 4:3, Vertical 4:5, Horizontal 5:4, Vertical 9:16, Horizontal 16:9, Horizontal 21:9
Seed	The starting dice roll. The same seed and settings give the same result every time; change it for a fresh variation, or leave it at -1 for a new random one.	-1 to 999999

Nano Banana Pro

Not the one for: No auto aspect ratio: the output snaps to the closest of ten fixed ratios, so an odd input shape does not come back as-is. 1K and 2K cost the same and 4K doubles the price; no mask input.

Behind the settings button:

Setting	What it does	Choices
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Match Input, Horizontal 21:9, Horizontal 16:9, Horizontal 3:2, Horizontal 4:3, Horizontal 5:4, Square 1:1, Vertical 4:5, Vertical 3:4, Vertical 2:3, Vertical 9:16
Resolution	How big and detailed the result is. Higher looks sharper but costs more and takes longer.	1K, 2K, 4K
Seed	The starting dice roll. The same seed and settings give the same result every time; change it for a fresh variation, or leave it at -1 for a new random one.	-1 to 999999

GPT Image 2

Not the one for: Not for region-locked inpainting: pixels outside its optional mask drift, which is why the Inpaint tool runs on FLUX Fill instead. The high tier is the expensive one.

Behind the settings button:

Setting	What it does	Choices
Quality	The overall finish level. Higher looks better but costs more and is slower.	Low, Medium, High
Aspect Ratio	The size and shape of the image. Bigger looks sharper but costs more and is slower.	Auto, Square, Horizontal 4:3, Horizontal 16:9, Vertical 3:4, Vertical 9:16
Size	Output resolution. Large (2K) renders the edit at a 2048 long-edge; 4K fills an 8.3 megapixel budget, so a wide edit comes back 3840 across and a square one 2880. Both cost more credits.	Standard, Large, 4K
Output Format	The file type: PNG keeps the best quality and supports transparency, JPG is a smaller file, WebP is small and modern.	PNG, JPEG, WebP

GPT 2.5 Flare

Not the one for: Caps at six reference images and changes only what the prompt names, so it is the wrong pick for a bigger reference set or for a full restyle that should reinterpret the whole frame.

Behind the settings button:

Setting	What it does	Choices
Quality	How hard the model works on the edit. Higher looks better, costs more and takes longer (5, 6, 12, 16 and 34 credits for a square picture at Standard size).	Low, Medium, High, Very High, Max
Aspect Ratio	The size and shape of the image. Bigger looks sharper but costs more and is slower.	Auto, Square, Horizontal 4:3, Horizontal 16:9, Vertical 3:4, Vertical 9:16
Size	Output resolution. Large (2K) renders the edit at a 2048 long-edge; 4K fills an 8.3 megapixel budget, so a wide edit comes back 3840 across and a square one 2880. Both cost more credits.	Standard, Large, 4K
Output Format	The file type: PNG keeps the best quality and supports transparency, JPG is a smaller file, WebP is small and modern.	PNG, JPEG, WebP

GPT 2.5 Sunburst

Not the one for: Noticeably slower than Flare for the same bill, so it is the wrong pick when turnaround matters more than fine detail, and it also caps at six reference images.

Behind the settings button:

Setting	What it does	Choices
Quality	How hard the model works on the edit. Higher looks better, costs more and takes longer (5, 6, 12, 16 and 34 credits for a square picture at Standard size).	Low, Medium, High, Very High, Max
Aspect Ratio	The size and shape of the image. Bigger looks sharper but costs more and is slower.	Auto, Square, Horizontal 4:3, Horizontal 16:9, Vertical 3:4, Vertical 9:16
Size	Output resolution. Large (2K) renders the edit at a 2048 long-edge; 4K fills an 8.3 megapixel budget, so a wide edit comes back 3840 across and a square one 2880. Both cost more credits.	Standard, Large, 4K
Output Format	The file type: PNG keeps the best quality and supports transparency, JPG is a smaller file, WebP is small and modern.	PNG, JPEG, WebP

Seedream 5 Pro

Not the one for: Output tops out at the 2K preset and there is no mask, quality ladder or seed here, so it is the wrong pick for a 4K deliverable or a strictly masked local fix.

Behind the settings button:

Setting	What it does	Choices
Image Size	The size and shape of the image. Bigger looks sharper but costs more and is slower.	Auto 2K, Auto 1K, Square small, Square large, Horizontal 16:9, Vertical 9:16, Horizontal 4:3, Vertical 3:4

Flux 2 Max

Not the one for: Costs about twice Pro per megapixel for the same request shape, so when the Pro result already reads right, Max is the pricier route. No mask input, and output above 2 MP is not recommended.

Behind the settings button:

Setting	What it does	Choices
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Match Input, Square 1:1, Horizontal 16:9, Vertical 9:16, Horizontal 4:3, Vertical 3:4, Horizontal 3:2, Vertical 2:3, Vertical 4:5, Horizontal 5:4
Resolution	How big and detailed the result is. Higher looks sharper but costs more and takes longer.	Match Input, 0.5 MP, 1 MP, 2 MP, 4 MP
Output Format	The file type: PNG keeps the best quality and supports transparency, JPG is a smaller file, WebP is small and modern.	PNG, WebP, JPG
Seed	The starting dice roll. The same seed and settings give the same result every time; change it for a fresh variation, or leave it at -1 for a new random one.	-1 to 999999
Safety Tolerance	How strict the content filter is. Lower is more cautious about blocking edgy content; higher lets more through.	1 to 5

Flux 2 Pro

Not the one for: No mask, so a change confined to one painted region belongs to inpainting, and it re-composes freely from references rather than holding a subject at its pixel position. Output above 2 MP is not recommended.

Behind the settings button:

Setting	What it does	Choices
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Match Input, Square 1:1, Horizontal 16:9, Vertical 9:16, Horizontal 4:3, Vertical 3:4, Horizontal 3:2, Vertical 2:3, Vertical 4:5, Horizontal 5:4
Resolution	How big and detailed the result is. Higher looks sharper but costs more and takes longer.	Match Input, 0.5 MP, 1 MP, 2 MP, 4 MP
Output Format	The file type: PNG keeps the best quality and supports transparency, JPG is a smaller file, WebP is small and modern.	PNG, WebP, JPG
Seed	The starting dice roll. The same seed and settings give the same result every time; change it for a fresh variation, or leave it at -1 for a new random one.	-1 to 999999
Safety Tolerance	How strict the content filter is. Lower is more cautious about blocking edgy content; higher lets more through.	1 to 5

GPT Image 1.5

Not the one for: Only three aspect ratios (1:1, 3:2, 2:3), no mask input and no batch count. The high tier is the expensive one; the image-edit pill opens it at medium and text-to-image at high.

Behind the settings button:

Setting	What it does	Choices
Quality	The overall finish level. Higher looks better but costs more and is slower.	Low, Medium, High
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Match Input, Square 1:1, Horizontal 3:2, Vertical 2:3
Input Fidelity	How faithfully it keeps the original’s fine detail while editing.	Low, High

P Image Edit

Not the one for: There is no quality or size setting, only an aspect ratio, so it is the wrong pick when the result has to be 2K or 4K or when fine detail and typography must hold.

Behind the settings button:

Setting	What it does	Choices
Mode	Picks the kind of edit, like relight, upscale, or turn-to-anime. Each mode does one specific job.	Default, Fusion, Extract Texture, Apply Texture, Relight, Multiple Angles, Upscale, Style Consistency, Subject Consistency, Anything to Real, Light Restoration, White to Scene, Add Characters, Next Scene, Scene Consistency, To Anime, To 3D Chibi, To Caricature, Photous, White Film to Rendering
Turbo	A faster, slightly lower-finish version.	On or off
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Match Input, Square 1:1, Horizontal 16:9, Vertical 9:16, Horizontal 4:3, Vertical 3:4
Seed	The starting dice roll. The same seed and settings give the same result every time; change it for a fresh variation, or leave it at -1 for a new random one.	-1 to 999999

Uni-1.1

Not the one for: Output size always follows the source at 2K and there is no aspect ratio control on an edit, so it cannot reframe a picture into a new shape; for a final pass the Max tier gives better output at a higher price.

Behind the settings button:

Setting	What it does	Choices
Style	Manga uses ink outlines and screentone shading.	Auto, Manga
Format	Output file format. Auto lets the model pick PNG or JPEG by content.	Auto, PNG, JPEG
Web Search	Let the model search the web for visual references before generating. Adds a few seconds.	On or off

Uni-1.1 Max

Not the one for: Costs more than the standard tier for the same controls, so it is the wrong pick for exploratory rounds or an edit that only needs a rough check; output still follows the source at 2K.

Behind the settings button:

Setting	What it does	Choices
Style	Manga uses ink outlines and screentone shading.	Auto, Manga
Format	Output file format. Auto lets the model pick PNG or JPEG by content.	Auto, PNG, JPEG
Web Search	Let the model search the web for visual references before generating. Adds a few seconds.	On or off

Qwen Image 3

Not the one for: Takes at most three images and a run lasts around two minutes, so it is the wrong pick for a four-image blend or for a quick rough pass.

Behind the settings button:

Setting	What it does	Choices
Image Size	The size and shape of the image. Bigger looks sharper but costs more and is slower.	Auto, Square large, Square small, Horizontal 16:9, Vertical 9:16, Horizontal 4:3, Vertical 3:4
Prompt Expansion	Quietly fleshes out a short prompt with extra detail before generating.	On or off

FLUX 3

Not the one for: A move is the unreliable edit: it runs on its own and sometimes redraws the picture around it. No seed. Not a restyle of the whole picture.

Behind the settings button:

Setting	What it does	Choices
Aspect Ratio	The shape of the result, like 1:1 square, 16:9 wide, or 9:16 tall. “Match input” copies the source shape.	Auto, Square 1:1, Vertical 2:3, Horizontal 3:2, Vertical 3:4, Horizontal 4:3, Vertical 4:5, Horizontal 5:4, Vertical 9:16, Horizontal 16:9, Horizontal 21:9
Resolution	How big and detailed the result is. Higher looks sharper but costs more and takes longer.	1K, 2K, 4K