心宝❤BananaV3

This is the "I need this to behave like a job queue, not a sampler" version of the pack's Gemini node. V3 adds two things the earlier versions don't have: OpenAI's GPT Image 2 as a first-class model option, and a submit-and-walk-away mode where generation happens in the pack's background task center instead of blocking your graph.

If V1 is "type a prompt, wait," V3 is "fire the request, keep editing the workflow, collect the image when the task center says it's done." For the batch e-commerce workflows this pack targets, that's the difference between babysitting a generation and actually getting other work done.

What changed vs V2

model_type now spans seven choices. The five Gemini models are still there, plus gpt-image-2 and gpt-image-2-oai - OpenAI's image model via the same reseller channel. That's the headline: one node, both Google and OpenAI image models, one key.

仅提交不等待 (submit only, don't wait), default off. With this on, the node submits the task to the pack's task center and returns immediately. Progress shows in the bottom-right "task center" panel (the same one the video nodes use), and the result comes back when the poller finishes. This is the async mode, and it's how you run a big batch without pinning ComfyUI's UI to a spinner. The trade-off is that your downstream nodes wait on the task's result rather than on the request.

GPT Image sizing controls. When a GPT model is selected, size offers 25 presets - auto, then 1K and 2K presets at 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9 and more - plus custom_width and custom_height (0–3840, step 16) that override size when both are above zero. And quality (low/medium/high, default medium) - note the tooltip: the non-OAI gpt-image-2 ignores quality, only the OAI build honors it.

Four outputs instead of three. You get images (single tensor), image_list (the same batch as an IMAGE list, useful when downstream nodes iterate per image), text, and failed_urls. The failed_urls output is the same rescue port as V2 - feed it to the 心宝❤失败链接汇总 aggregator.

What's the same

All the Gemini behavior carries over: prompt, banana_api_key (blank inherits from config.ini), aspect_ratio, seed, top_p, image_size, five reference images, binding_context, 联网搜索, 启用工作流并发, and the >5M限制长边 upload policy. One difference worth noting: V3 only offers the 心宝❤新渠道 route - the older Hong Kong/US/CF channel options are gone here, because this node targets the newer channel's async task system.

Installing and running

Same pack, same steps as the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/98624017/comfyui-banana-li-linux

Restart ComfyUI. Dependencies are the pack's usual: requests, opencv-contrib-python, scikit-image, segment-anything, blend_modes, timm, Pillow, psd-tools.

Where people trip up: they enable submit-only mode, watch the graph "finish" instantly, and then wonder where the image is - it's in the task center, still rendering. Check there before assuming the run failed. And remember the billing reality from the KB's closed-source models analysis: GPT Image and Gemini Pro are both metered per image, and async mode makes it easier to burn through a balance without noticing, because nothing is visibly waiting. Treat the task center like a meter, not a fire-and-forget convenience.