Nano Banana 2.1 · Text to Image

Why you'd call a cloud model for a text-to-image job

Because the thing you want has letters in it.

Nano Banana is Google's Gemini-native image family, and 2.1 is the current "newest, sharpest" cut of it - the pitch being legible text and 4K-class output. If you've spent an evening fighting FLUX or Qwen-Image over a poster headline, a shop sign, or a four-word brand mark, you already know the pain this node sells a fix for. Local models have closed a lot of ground, but dense typography at high resolution is the one place where the closed frontier still visibly wins.

The honest counterweight: this is an API node. Your prompt leaves the machine, Google's safety filter is the final authority on what you get back, and nothing about your workflow makes it cheaper over time. For one hero image a week, that's a fine trade. For a nightly batch of 500 stock-ish images, run the numbers before you fall in love.

The mechanics

The node posts your prompt and settings, takes a quote, submits the task under an idempotency key, polls until it settles, then downloads the returned files and decodes them into a normal IMAGE batch. Multiple images come back as a batch, so a Save Image node writes them all. The pack installs nothing - its pyproject.toml lists no dependencies, and it uses the aiohttp, torch and Pillow that ComfyUI already ships.

The flags you won't find on a local sampler:

• enable_web_search - the model checks current facts on the web before drawing. Off by default.
• enable_image_search - it looks up reference pictures instead of inventing them. Also off.

Both are latency for accuracy. Turn one on for "the 2026 lineup", "this week's scoreboard", "our actual product", and leave them off for anything a photograph could settle.

Inputs

• prompt - describe the image you want. With this family, be specific about the words that should appear and put them in quotes; it renders literal text well enough that ambiguity is your fault, not the model's.
• aspect_ratio - default 16:9 here, which is a sensible canvas for a poster or a slide. Portrait for phone content.
• resolution - the output size tier; 1k default, higher tiers cost more. Start at 1k while you're still writing the prompt, bump it when the composition is locked.
• enable_web_search, enable_image_search - as above.
• output_format - png by default. Keep it for anything with text in it; you want zero JPEG artifacts around letterforms.

Outputs: images (IMAGE) and task id (STRING). The task id is what you quote when a run comes back wrong, and what you reference in the PopNow API console.

Install

comfy node install popnow-comfyui

or, in ComfyUI Manager, search PopNow and hit Install. Restart, generate a key at popnow.ai/api/console/keys, paste it under Settings > PopNow (or export POPNOW_API_KEY for a headless box), then restart a second time so the node list loads from your account's catalog. Delete custom_nodes/popnow_api.py if you installed the old single-file version - it's the same node type and two copies cause nothing but confusion.

No GPU needed, no weights downloaded, no Python wrestling. You need a recent ComfyUI though: the pack is built on the V3 node API.

Costs and failures

About 7.2 credits per image at the default tier, quoted before every run, with the final charge printed on the node when it finishes. Set Max credits per run in the PopNow settings so a 4K sweep can't quietly eat your balance.

PopNow 401 is a bad key. 40201 is an empty wallet. 400 names the field it didn't like - usually resolution against aspect ratio, or a string where a number belongs. 40004 means the combination isn't servable. 40310 wants your email verified. content_rejected means the filter declined; the credits come back, your time doesn't.

Two behaviours worth remembering: seeds aren't a widget on this node, so every queue is a fresh billed image, and cancelling only stops you waiting - the task keeps running on PopNow and is charged if it lands.

This pack is new and the vendor behind it is essentially undiscussed in the usual communities, so treat it as an unproven tool: keep the key in the environment, keep credit purchases small, and read the source if you plan to depend on it.