Triple API Generate (NanoBanana + Doubao + HiDream)

Type one prompt, get three images back from three different hosted services at once. That's the pitch of Triple API Generate, the flagship node of the AlexXia007/AIYang_TripleAPI pack - it fires NanoBanana, Doubao, and the 即梦 (Jimeng) visual API in parallel and hands you each result as a separate image. It's a quality-comparison rig, a hedge against one flaky API, and a fast way to spend credits on all three platforms simultaneously. No local model, no GPU: you supply the API keys and the services do the work.

The naming is worth a moment

Before you trust the labels, know what you're actually calling. "HiDream" in this node is not the Beijing HiDream.ai that ships MIT open weights - it's 即梦, ByteDance's image API on Volcano Engine (the default server is visual.volcengineapi.com, the Seedream family). And "NanoBanana" is the third-party hosted API at api.nanobananaapi.ai, not Google's Nano Banana. So "three platforms" is really two ByteDance/Volcano properties plus one outside service. ByteDance keeps Seedream API-only by design, which is exactly why this node exists: it's the cloud path for models you can't run locally.

How it works

Under the hood it's a ThreadPoolExecutor with three workers, so the calls genuinely run in parallel - each with its own retry loop (max_retries, retry_interval) and its own polling for the async generation jobs. Fault isolation is the whole point: if Doubao errors, NanoBanana and Jimeng still come back. A failed or disabled platform doesn't kill the run; it returns a white placeholder tensor (or your fallback_image_url if you set one) and records what happened in status_info. The *_image_urls outputs follow the README's convention: a real URL list on success, ["api_error"] on failure, ["api_disabled"] when you left the key empty or the toggle off.

The inputs and outputs that matter

Setup is really just three things:

• prompt - the one input all three platforms share.
• The key/server pairs - nanobanana_api_key, doubao_api_key, hidream_req_key, each with a *_server_url default that you can usually leave alone. Fill only the keys you own, and use the enable_nanobanana / enable_doubao / enable_hidream toggles to switch off the rest - otherwise a keyless platform just burns a slot on a white image.
• seed - shared across all three for reproducible comparisons.

The optional bank is where the per-platform tuning lives: image_urls_input for the img2img modes (Doubao's img2img uses the first URL, multi_img_fusion uses them all, Jimeng takes up to 10), doubao_model (default doubao-seedream-4-0-250828), doubao_size (1K/2K/4K), hidream_type (including shopImg2img and vton), and hidream_scale for edit strength. Leave them until you know you need them.

Outputs: nanobanana_image, doubao_image, and hidream_image (all IMAGE) - wire each into a Preview Image or Save Image node. The *_task_id, *_response, and status_info strings go to ShowText nodes when you're diagnosing a failure, and *_image_urls are handy if you want to hand the result off to another system without re-downloading.

Install

Same as any node from this pack - no models, no heavy deps (Pillow/numpy/torch, all already in ComfyUI):

cd ComfyUI/custom_nodes
git clone https://github.com/AlexXia007/AIYang_TripleAPI.git
cd AIYang_TripleAPI
pip install -r requirements.txt

Restart ComfyUI and it appears as Triple API Generate (NanoBanana + Doubao + HiDream) under the AIYang_TripleAPI category. (The README tells you to cd AIYang_comfyui_tripapi after cloning - that directory doesn't exist; it's AIYang_TripleAPI.)

Common issues

• Three keys is the real setup cost. You need accounts at nanobananaapi.ai, the Volcano Engine console for Doubao, and 即梦's developer portal. There's no way around it; the node is only as useful as the keys you paste in.
• White images on a slot. That's a disabled platform or an empty key. Disable what you don't use, or you'll spend a while wondering why one output is a blank square.
• Slow runs. Everything is async and polls on your timers; bump poll_timeout (default 240s) if generations time out.
• It's real money per image on every platform, run in parallel by design. A failed comparison costs you three API calls, not one. The README is also Chinese-only and rot-prone - it documents an img2img sibling node that isn't registered in the shipped code.