arkennemasis fal · Nano Banana 2 · $0.06-0.16/image

Nano Banana 2 is the Gemini 3.1 Flash generation of Google's image line - the speed-quality hybrid that landed after the Pro model. Base rate on fal is $0.08 an image, and the interesting part isn't the pixels. It's that this endpoint can search the web mid-request and think before it draws.

That matters more than it sounds. A local model generates what it memorised. This one can go check what a specific product's packaging looks like now, then draw it. For anything time-sensitive or brand-specific, that's a capability nothing you can download has.

What the settings do, priced honestly

The price isn't flat, and the node's live badge follows your widgets:

• resolution - 0.5K, 1K (default), 2K, 4K. Billed at 0.75× for 0.5K, 1× at 1K, 1.5× at 2K and 2× at 4K. So a 4K image is sixteen cents, not eight.
• enable_web_search - off by default, +$0.015 per request when on. A rounding error next to the image itself.
• thinking_level - (not set) by default; minimal or high turns on model thinking, which adds +$0.002 at high. Two-tenths of a cent to have the model reason about the composition first is the best-value toggle on this node.
• num_images - 1 to 4, each billed, with every multiplier applied per image.

Then the composition controls. aspect_ratio defaults to auto - the model reads the prompt and picks - and this family supports genuinely extreme ratios: 4:1, 1:4, 8:1 and 1:8 alongside the usual set. An 8:1 banner is something you cannot get out of a local SDXL workflow without a lot of stitching. system_prompt is a persona/style instruction applied across the request - a place to put "always add film grain, never add text" once instead of in every prompt. safety_tolerance (1–6, default 4) is the moderation dial, and per the KB this line is aggressively filtered: the January 2026 update tightened restrictions on famous IP and celebrity likeness. seed takes -1 for model's choice. output_format is png, jpeg or webp. limit_generations is on by default and disregards in-prompt requests for multiple variations - the usual cause of "I asked for four and got one".

Outputs

images - a batchable IMAGE output, which is what you'd feed a local upscaler, a ControlNet preprocessor, or an open video model's image input. image_paths - files under output/fal/fal-ai_nano-banana-2/. description - the model's own text about the result, ideal as a caption or as a handoff into an LLM node. info - JSON with the request id, the cost estimate, the saved files and fal's complete answer.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart ComfyUI, or search arkennemasis in ComfyUI-Manager. requirements.txt is a single line - httpx, for the pack's Codex nodes - because the fal provider talks to fal.ai with Python's standard library. Then:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

in the .env next to run_nvidia_gpu.bat. The pack deliberately ships no key input on any node, since widget values are written into workflow JSON verbatim.

Where people get burned

• The bill is bigger than the badge suggested. Resolution multipliers and num_images compound: four 4K images is 64 cents, not 32. The node's own price badge tracks this as you change widgets, so look at it.
• You asked for four images and got one. limit_generations, on by default.
• auto reframed your scene. Convenient until it isn't. Pin the ratio when shape matters - and note this family's extreme ratios are also useful for cheating a "wide establishing shot" out of a stills model.
• The web search gave you an old logo. Search is a tool, not a truth. The model still draws what it decides.
• It's a $0.08 toll on every experiment. The cheap first pass is base Nano Banana at $0.0398; come here when you need resolution, research or thinking.