Zhenzhen Image G v2.5 Official 生成/编辑（2 合 1）

Eleven required inputs. That's how you know this is the "real" node in the pair: Zhenzhen_Image_G25_Official is the 2-in-1 flagship of the Image G v2.5 line in ComfyUI_Seedance, where the LowPrice node next to it gives you four fields and a fixed ratio list. Here you get two models, an exact pixel size, output format, quality tiers, and up to four images per call.

It still doesn't run anything locally. This is an HTTP call to a paid gateway dressed as a sampler node, which is the whole trade: no VRAM, no checkpoint, per-call billing, and your prompt plus every reference image leaves your machine.

The two models, and which one you want

model is the first field and it matters more than anything else on the node:

• zhenzhen-image-g-v2.5-flare - the default. The author's own tooltip says it's for everyday creation and fast iteration.
• zhenzhen-image-g-v2.5-sunburst - "focuses on precise editing," same tooltip.

Read that as: start on Flare, switch to Sunburst when you're doing reference-image surgery and the edit keeps bleeding into parts you wanted left alone. Both share an identical input contract, so the switch costs you nothing but a queue.

Why you'd use a cloud node for this

Because this is the job you'd otherwise hand to Qwen-Image-Edit or Flux Kontext locally, which means a GPU, a tuned stack and a few model downloads. The honest counter-argument floats around every node in this category - "local and not through an api? No? Nvm then" was the loudest reaction to Seedance arriving in ComfyUI at all. Fair. Also the position of someone who already has the GPU. If you don't, or you want to know what 16-reference editing feels like before committing to a local pipeline, this is the shortcut.

What this node has over most cloud image nodes is control: sixteen reference slots, exact pixels, PNG/JPEG/WebP, transparency, quality tiers, a moderation level. Whether that's worth a per-image price is your call - and it's a third-party gateway (api.seedance.nz) with affiliate links on the README, not Comfy's official Partner Nodes. Third-party Seedance hosts are a known grey market, so don't assume the model name on the wire is the model you're imagining.

Inputs worth setting

Required, in the order they matter:

• prompt - required. It's an editing instruction when references are connected.
• size - the interesting one: eighteen options including auto, the usual ratios (1:1, 3:2, 2:3, 4:3, 3:4, 5:4, 4:5, 16:9, 9:16, 2:1, 1:2, 21:9, 9:21, 3:1, 1:3), plus preserve_reference and custom.
• custom_size - only used when size is custom, formatted WIDTHxHEIGHT, e.g. 1536x864.
• resolution - 1k/2k/4k. Per the tooltip, the API ignores it once you've given an exact pixel size.
• n - 1 to 4 images, returned as one IMAGE batch, not four outputs.
• quality - auto, low, medium, high, xhigh, max.
• output_format - png, jpeg or webp, with output_compression for the last two only.
• background - auto, transparent, opaque. Transparent needs PNG or WebP; transparent JPEG is rejected before submission.
• moderation - low or auto. A documented level on this host, not an off switch.

Optional: image1 … image16. Connecting any reference silently switches the request into editing mode. Plus api_config, skip_error for a placeholder instead of a broken workflow, and cache-only seed.

preserve_reference is the sleeper feature: it omits size from the request entirely so the API keeps your reference's aspect ratio - exactly what you want editing a photo whose framing shouldn't change.

The builder also validates what the docs demand, so you find out at queue time rather than five minutes later: custom_size needs multiples of 16 on both sides, neither over 3840, an aspect ratio no worse than 3:1, and total pixels between 655360 and 8294400.

The trap: transparent background won't survive

Ask for background=transparent and PNG, and yes, the API produces an alpha channel. Then the node downloads the result and converts it to plain RGB before handing it to ComfyUI. The alpha is gone. You get a flattened image and no mask output on this node, so if your workflow depends on cutting the subject out, this is not the node - the pack's layer-decomposition nodes are the ones that return IMAGE plus MASK.

Outputs

image is the batch of results as a ComfyUI tensor (one to four images), straight into Save Image or an upscaler. image_url is the temporary link, task_id the remote job ID, response the raw JSON. Save results you care about; those links expire.

Install

Manager: search ComfyUI Seedance (registry publisher t8star, node ID seedance), or:

comfy node install seedance

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
cd ..
python -m pip install -r custom_nodes/ComfyUI_Seedance/requirements.txt

The only dependency is requests. No model files, no FFmpeg needed for this node, no torch changes. Restart and look under the Seedance category. Then connect Seedance API Config (base_url https://api.seedance.nz plus your key), or set SEEDANCE_API_KEY, or write config/.env next to the pack.

When it goes wrong

• "custom size must use WIDTHxHEIGHT", or a pixels/ratio complaint - fix custom_size; the validator is strict on purpose.
• "transparent background requires png or webp" - you picked JPEG with transparency.
• n=4 and Save Image writes four files - that's ComfyUI batching, not a bug. Index the batch if you wanted one.
• A blank or oversized prompt halts the run before submission; the prompt runs 1 to 5000 characters.
• Refusals on faces - closed hosts filter at the model and there's nothing local to patch. Dial moderation if your route allows it, but don't expect guarantees.
• SSL certificate errors on Windows portable - python -m pip install -U requests certifi in ComfyUI's own Python, then consider SEEDANCE_CA_BUNDLE.