并发提交｜Qwen Image 3.0 / Pro 图像生成/编辑（8 合 1）

Qwen Image 3.0 is Alibaba's image model, and through this pack you get eight flavors of it - standard and Pro, domestic and global-, each with a text-to-image and an image-to-image variant. This node is the concurrent-submit wrapper around that 8-in-1, and it's a strong one for iteration work because each job can also request up to six images at once (n). Stack a few of these into the 30-slot image receiver and you can effectively flood a canvas with candidates in one run.

The pattern is pack-standard: set your parameters, the node validates them, then the job runs on a background image thread pool and returns a SEEDANCE_IMAGE_FUTURE. Feed that into SeedanceConcurrent_Image_Await, which waits on up to thirty image futures and outputs them in slot order. If you're hunting a look, this is the node that lets you fire ten Qwen variants and pick.

The inputs that matter

• model - the 8-way dropdown. I2I variants need reference images; T2I ignores them.
• prompt - 5 to 2000 characters, required.
• negative_prompt - optional, genuinely supported here.
• prompt_extend - on by default; lets the upstream rewrite/expand your prompt. Turn it off if you want raw control.
• sizing_mode - the one that trips people. auto sends no size at all; ratio sends metadata.ratio + resolution (1k/2k); custom_size sends a top-level size like 1024*1024. The three modes are mutually exclusive request shapes, so the resolution/ratio/custom_size controls only apply in their own mode.
• n - 1 to 6 images per job.
• seed - -1 omits the seed from the request; a non-negative value gets forwarded. fixed on the ComfyUI side also reuses cached results.
• image1…image3 - the I2I reference bank; at least one needed for an edit, max three.

Plus api_config and skip_error (placeholder image on failure rather than killing the batch).

How it runs

Standard wrapper machinery, with the front end hiding whichever size controls don't apply to the current sizing_mode. Reference images are uploaded automatically, and the worker does submit → poll → download, returning a normal IMAGE tensor. The image pool defaults to 30 workers (SEEDANCE_IMAGE_CONCURRENCY to tune).

Installing

# ComfyUI Manager: search "ComfyUI Seedance" → install → restart.
# or:
comfy node install seedance
# or:
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

Only requests; no weights. API key via Seedance API Config, SEEDANCE_API_KEY, or config/.env.

Where people trip

• Sizing confusion - if you pick auto, your carefully-set custom_size does nothing, and vice versa. The node's tooltips spell out which fields fire in which mode; trust them.
• I2I with no reference - validation bounces it. T2I and I2I are split by the model choice, not by whether an image is attached.
• n > 1 on a big batch - remember each slot's job multiplies: 10 slots × n=6 is 60 metered image calls. The math gets real fast.

Pack-wide fine print: prompts and reference images leave the machine, every image is metered, and links expire - save what you keep.