Jimeng Image 4.0

Let's be upfront about what this node is: it's a thin wrapper around ByteDance's Jimeng (即梦) 4.0 image model, served through the Volcengine API. No local GPU, no model download - you're renting generation from a Chinese cloud service that has, like almost everything ByteDance ships in this space, stayed resolutely closed-source. You install it because you want Jimeng 4.0's output inside your ComfyUI graph without tabbing out to a web app, or because you want to chain its results into the rest of a local pipeline - upscale the output, batch it, compare it against your local checkpoints.

The catch, and it's a real one: you're paying for this. Every run is a live API call, and the node deliberately disables ComfyUI's caching (IS_CHANGED always returns a fresh timestamp), which means every graph execution fires a fresh, billable request. Accidentally rerun a prompt that has the node sitting upstream and you've just spent credits on images you didn't want. Treat it like a printer: think before you hit queue.

The three-key problem

Most API nodes need one key. This one needs three, and the third is the weird one. You need a Volcengine access_key and secret_key (grab them at volcengine.com), but you also need a PicGo API key - a separate signup at picgo.net. Why? Because the Jimeng 4.0 API won't accept base64-encoded images, only public URLs. So the moment you connect an input image, the node uploads it to PicGo's image hosting and passes back a URL. If you feed it an image without a picgo_api_key, it hard-errors. Non-Chinese users should also brace themselves: Volcengine's console, billing, and docs are heavily CN-oriented, and that friction is real community signal, not paranoia.

How it works

Under the hood it uses Volcengine's official Python SDK, which the pack vendors inside lib/volcengine - no pip install needed. It submits an async task (request key jimeng_t2i_v40) and then polls for the result every poll_interval seconds until timeout (defaults: 5s and 300s) gives up. The dynamic input slots are the part to know: connect an image and a second image_2 slot appears, up to ten of them, for image editing and multi-image composition rather than plain text-to-image. Output is an images list (IMAGE), ready for any downstream node.

The inputs that matter

• prompt - the text prompt. Defaults to a Chinese phrase ("一幅美丽的风景画"), so type over it.
• force_single and scale - control whether you get one image or a grid, and the guidance scale.
• size_mode - auto, custom_size (total pixels, default 4MP), or custom_dimensions (explicit width/height, 512–4096).
• min_ratio / max_ratio - aspect-ratio bounds (default 0.33 / 3.0).
• add_watermark plus the watermark_position / language / opacity / text options - ByteDance can stamp a watermark for you if you toggle it.
• timeout and poll_interval - tune how long the node waits on the async job.

Installing it

Same as every node in the WBLESS pack:

cd ComfyUI/custom_nodes
git clone https://github.com/LaoMaoBoss/ComfyUI-WBLESS.git

then restart, or use ComfyUI Manager and search "ComfyUI-WBLESS". No extra pip installs - the Volcengine SDK ships in lib/. If you ever see "火山引擎SDK未找到" in the console, a restart usually fixes it since the local SDK path is registered at import; worst case, pip install volcengine. Then paste your three keys and accept that you've turned ComfyUI into a very fancy API client.