Nano Banana 2 · Edit

The one you'll actually use most days

Nano Banana 2 is the speed-quality hybrid of Google's Gemini image line - the February 2026 release that sits under the Pro flagship on quality and over it on cost-per-image. In the PopNow catalog it's about 5.4 credits an image against 2.1's 7.2, and for most editing work the difference doesn't show. Composite a product onto a new background, change the light, swap the jacket, remove the logo, extend a set 20% wider - this node handles all of it, up to 4K, with up to fourteen references.

Save the pricier nodes for the jobs that need them: dense multi-language text, diagrams, and anything where the model should reason about layout first. For everything else, this is the default.

How the edit actually happens

There's no mask input and no inpainting model. You're describing a change in language against one or more source images, and the model does the rest. That's the practical difference from the local workflow you may be used to: where Qwen-Image-Edit or a Kontext-style pipeline hands you pixel-accurate control and full privacy, this hands you a sentence and takes your pictures to a server.

Under the hood the node encodes your image batch - PNG, or JPEG at high quality if a frame is over 10 MiB - uploads it, caches that upload for twelve hours on the content hash so re-prompting doesn't resend it, quotes the run, submits it under an idempotency key, polls, downloads, decodes to IMAGE. Plain HTTP client, no local model, no VRAM.

Fields

• prompt - the author's tooltip says it better than I can: "Say what to change and what must stay the same." Both halves are load-bearing. Edits that drift come from prompts that only state the change.
• image_urls - an autogrow socket group, one to fourteen images, in the order your prompt refers to them. The socket names are positional (image_0, image_1…) because the group label isn't drawn on the node. Reference them out loud: "in the second image".
• aspect_ratio - auto by default, preserving the source shape. Setting it explicitly is a reframe request.
• resolution - the size tier, 1k by default. Higher tiers cost more per image.
• enable_web_search - look up current facts before drawing.
• enable_image_search - look up reference images before drawing.

Note what's not here compared with the 2.1 node: no output_format widget. You take what the tier gives you.

Outputs: images (IMAGE) and task id (STRING).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart, add a key from popnow.ai/api/console/keys under Settings > PopNow (or set POPNOW_API_KEY), restart again so the catalog loads. No dependencies, no downloads, no GPU. If you're on an older ComfyUI, update first - the pack needs the V3 node API.

Field notes

Ordering is the number one source of bad results. Fourteen references in a jumble and a vague sentence is how you get a collage nobody asked for. Three well-described references beat twelve.

PopNow 400 naming a field is a value problem, not a node problem; 40004 is a combination the model won't serve; 40201 is credits; 40310 is email verification. Declined images come back as content_rejected and are refunded. Every run is quoted first, and Max credits per run in settings is a real ceiling - it stops the run before anything is held.

Because it's a per-image charge with no seed, each Queue is a new billed attempt. Get the sentence right at 1k, then spend on the higher tier once.

And the standing caveat for this whole pack: PopNow is young and effectively absent from community discussion, so there's no accumulated wisdom to borrow. The code is MIT and small enough to read; the key stays in settings or the environment so a shared workflow can't leak it. Treat the first month as a trial.