SF WaveSpeed Nano Banana Pro Text to Image

You can't download Nano Banana Pro - it's Gemini 3 Pro Image, closed, living on Google's infrastructure. But you can rent it per image, and that's exactly what SF WaveSpeed Nano Banana Pro Text to Image does: a prompt goes in, a real IMAGE tensor comes out, and your WaveSpeed account gets billed by the call. For text-to-image, the appeal is the same as for editing - this is a model with a "thinking" pass before it draws, native multi-language text rendering, and genuinely strong prompt adherence, and you get it without any Gemini credentials, just the key you already use for the rest of the WaveSpeed nodes.

How it works

The node sends your prompt, aspect ratio, and resolution to WaveSpeed's google/nano-banana-pro/text-to-image endpoint and decodes the returned image into a ComfyUI tensor. Pricing is printed right in the tooltip - 1k and 2k at $0.14, 4k at $0.24 - which makes the economics obvious: iterate at 1k, spend the big resolution only on finals. The output_format defaults to PNG, which matters if you'll be cutting the result out or stacking it with transparency. Output is output_image (IMAGE), so everything downstream behaves like a local generation.

The inputs that matter

• client - from SF WaveSpeed Client, wired first. No client, no image.
• prompt - the generation prompt. Nano Banana Pro rewards specificity and composition language; the thinking pass is only as good as the brief you give it.
• aspect_ratio - default 1:1; set it to match your target canvas.
• resolution - 1k / 2k / 4k. Start at 1k.
• output_format - PNG (default) or JPEG.
• enable_sync_mode - default true; leave it.

Output: output_image (IMAGE).

Installing it

In the SF ComfyUI Nodes pack from Stillfront. ComfyUI Manager → "SF ComfyUI Nodes" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Stillfront/comfyui-sf-nodes.git
cd comfyui-sf-nodes
pip install -r requirements.txt

Plus a WaveSpeed API key through SF WaveSpeed Client. No models download; the generation happens on WaveSpeed's side.

Common issues

The failures here are mostly about expectations rather than wiring. This is a closed model with Google's safety policy baked in - the KB's honest summary is that the Nano Banana line is powerful but "unimaginative" when the filter decides to engage, so don't burn an evening fighting refusals. And while the per-call price is low, it's not zero: at 4K the cost per image is real, so budget for iteration. One more thing - because output is a single tensor, if you're used to diffusion models spitting out a batch of four to pick from, this node gives you one; run it a few times or wire up a rerun loop instead of assuming a batch.