心宝❤BananaV2

If you've already got the base 心宝❤Banana node working, V2 is the same engine with three things bolted on that matter for real workflows: an internet-search toggle, a workflow-concurrency mode, and a failed_urls output that lets you rescue the images Gemini claims it made but never handed you.

It's the same Gemini image API under the hood - same five models, same batch, aspect ratio, seed, and up-to-five reference images. In fact every input and output on V1 is here. The differences are all in the direction of "production batch work," which is the point of this pack: Xinbao built it for e-commerce detail pages where you're generating dozens of images per session and can't afford a single failure to nuke the whole run.

What's new

启用工作流并发 (workflow concurrency), default off. This is the big one. Normally every Banana node fires its API call one at a time in graph order. Flip this on and multiple BananaV2 and video nodes in the same workflow fire concurrently. The trade-off the author made: in concurrency mode, a single node's failure doesn't kill the graph - that node's output just blocks and the error text appears in its text port. You get resilience at the cost of determinism. It's off by default to keep resource usage conservative; you turn it on when the session is long.

联网搜索 (web search), default off. The model can search the web to assist generation. The tooltip is refreshingly honest: image models may just ignore this tool, so treat it as a "sometimes helps, never hurts" switch rather than a feature.

>5M限制长边 (over-5MB long-edge policy). Uploaded reference images bigger than 5MB get rejected by the API, so this controls what happens before upload. 禁用 (off) just re-encodes to JPEG q90 - keeps resolution, shrinks the file. The 2048 / 3072 options force-resize the long edge down to that pixel count instead. If your product photos are huge, this is the knob that stops upload failures.

And the reason this node has three outputs where V1 has two: failed_urls. When a batch image comes back as a URL that then fails to download locally, the URL goes here instead of being silently lost. That output feeds the pack's companion node, 心宝❤失败链接汇总 (BananaFailedUrlAggregator), which collects failed URLs across many V2 nodes and retries them.

The rest, briefly

images and text behave exactly like V1 - text carries the error message when things go wrong. The binding_context input is there for the local-crop chain. And everything about 线路 (route), 绕过代理, 禁用SSL验证, top_p, and image_size carries over unchanged, as does the default of gemini-3-pro-image-preview and the batch-1-4-for-stability advice.

Installing and gotchas

Same pack, same install. Manager search "comfyui-banana-li", or clone the Linux repo:

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

Restart, then set your key once in config.ini ([gemini] section) or leave banana_api_key blank on the node to inherit it.

Where people get burned: they flip concurrency on, expect a 2x speedup, and are confused when a node mid-graph quietly blocks. That's the designed behavior - concurrency is a fire-and-forget mode, and you check the text outputs for what actually failed. And remember the upstream quirk the README hammers on: high batch (especially 8) can return "success" while billing you nothing back. V2's failed_urls output is the pack's answer to that - wire it to the aggregator rather than just re-running the whole batch.