Seedream 4.0 · Edit

Most "edit" nodes give you an image back at whatever size the model felt like. Seedream 4.0's edit node takes a size widget with 16 hand-picked resolutions on it, from 1024×1024 up to 4096×2304, and delivers exactly that. If you're compositing against a fixed layout, or building a batch that has to line up, that alone justifies the node. At 4.5 credits (~$0.045) per image it's also cheap enough to iterate on.

It accepts multiple source images too - the model's headline is editing with up to ten references at once, so a character, a product, a background and a style board can all go in and the prompt explains what each one contributes.

What it is under the hood

A node in the popnow-comfyui pack calling ByteDance's bytedance/seedream-4.0/edit on PopNow's servers. The image_urls input is an autogrow socket group - connect a Load Image, another appears, up to the model's limit. On queue, the images are uploaded (cached by content hash for 12 hours, so reruns don't re-send them), the request goes in, the results come back and get decoded into a ComfyUI IMAGE batch.

Seedream is a closed model, so this is the only way to run it inside ComfyUI. Costs are per image and quoted before the run; a failed run is refunded.

Inputs worth your attention

• prompt - describe the image you want to end up with from the sources: subject, composition, style, lighting. Not "change the shirt colour" - more like "the same woman, now in a linen shirt, same window light, same framing". Models of this class respond to whole-picture instructions better than to terse diffs.
• image_urls - the source images, and the tooltip's "that you are authorised to process" is doing legal work, not technical work: you're sending someone else's picture to a third party.
• size - width*height, and it's a dropdown, not a slider. Options run from 10241024 through square, landscape and portrait variants: 20482048 is the default, 23041728 and 17282304 for 4:3 both ways, 28481600 / 16002848 for wide, 31361344 for ultrawide, 30723072 and up to 4096*2304 at the top. Pick the one matching your delivery rather than letting a resample undo the good work.

Outputs

images - a batch, so multiple returned images all come through - and task id, the string for PopNow's API console logs or support. Wire images into Save Image, or into a local upscaler or detailer if you want to push further with your own models.

Note what's not here: no seed, no negative prompt, no CFG. This is a hosted model with a small surface area, and none of the local sampling vocabulary applies. If you're used to tuning a sampler, your only levers are the prompt, the sources and the size.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Or search PopNow in ComfyUI Manager and click Install. Then grab a key from popnow.ai/api/console/keys, paste it into Settings > PopNow (or set POPNOW_API_KEY before launching), and restart a second time - the node list is rebuilt from your account's catalog on startup, which is why a fresh model can seem "missing" until you restart.

No pip install needed. The pack's dependency list is empty and it uses only what ComfyUI already ships.

Where people trip

• PopNow 40201 - out of credits. At 4.5 per image you'll notice this quickly in a batch.
• PopNow 400 naming a field - usually too many source images, or a size you assumed was free-form. It's a fixed list; type a custom resolution and it'll be rejected.
• PopNow 40310 - verify your email address. Unexpected, and it does block runs.
• content_rejected - the model declined. Refunded, so reword rather than retry identically.
• Cancelling stops your wait, not the job. It keeps running on PopNow, bills if it succeeds, and the image is in your API console.

If you've been doing reference-based editing locally with IP-Adapter or a Qwen-Image-Edit workflow, this is the same idea with ten input slots and none of the setup. It won't give you a LoRA's consistency across a hundred images, but for one shot, one product, or one afternoon of variations, it's the shorter road.