Seedream 5.0 Flash · Edit

The most useful thing on this node isn't the editing. It's background: set it to transparent with output_format: png and the returned image has a real alpha channel - no rembg pass, no BiRefNet, no matting node, no fiddling with edge refinement on flyaway hair. Given how much of the ComfyUI ecosystem's background-removal conversation is about tolerating whatever the matting model does badly, a model that just outputs the cutout is worth knowing about.

4.35 credits (~$0.044) per image. Ten reference inputs. Three resolution tiers at one price. This is the workhorse of the Seedream 5.0 family.

How it works

PopNow's node for bytedance/seedream-5.0-flash/edit. image_urls is an autogrow socket group: connect one Load Image and another socket appears, up to ten. Your prompt describes the image you want back, not the delta, and explains what each reference is for. Images are uploaded on queue (content-hash cached for 12 hours), generated server-side, and returned decoded as an IMAGE batch.

The pack also adds an alpha output on this node specifically, because the model's schema has a background property - that's the pack's rule for when to expose a mask, and it's a good one to know: if you see alpha, something this model generates may be partially transparent.

The inputs that matter

• image_urls - one to ten references. Put your prompt in terms of the slots: "person from image 1, jacket from image 2, the lighting of image 3". It reads better than hoping.
• prompt - describe the target image, and say what each reference contributes.
• aspect_ratio - auto (default, keeps the first reference's shape) plus the usual ladder: 1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3, 4:5, 5:4, 21:9, 9:21, 2:1, 1:2, 3:1, 1:3. auto is right most of the time when you're editing an existing picture.
• resolution - 1k, 1.5k, 2k, and the tooltip is explicit that all three tiers cost the same. There's no reason to run 1k unless you want it fast. This is not the node to save credits on.
• output_format - jpeg or png. png is required for the transparent path.
• background - opaque or transparent. Transparent needs png output and, per the tooltip, a single PNG input with alpha.

Outputs and the mask gotcha

images, alpha (a MASK, 1 = opaque), task id.

The pack decodes transparency honestly: for transparent results, the RGB it hands you is already premultiplied by the alpha. That's what makes a transparent background look correct in a preview instead of showing garbage colour fringe. It also means you shouldn't multiply by the mask again downstream, which is the mistake people make when they wire alpha into Image Composite Masked and wonder why their cutout turned to mud. Take images as-is, use alpha for compositing or for masking.

If you just want the cutout, ignore alpha entirely and save the images batch as PNG - the transparency survives.

Install

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

Or Manager → search PopNow → Install. Then key it: popnow.ai/api/console/keys → Settings > PopNow or POPNOW_API_KEY, then restart once more so the node list rebuilds from your account. No pip dependencies - the pack deliberately declares none.

Troubleshooting

• alpha is all ones - background wasn't set to transparent, or output_format was jpeg. Both need to be right.
• PopNow 40004 - the transparent path has constraints; the tooltip's "one PNG input with alpha" is the one people skip.
• PopNow 40201 - credits. Flat per image, so batching ten variants is 43.5 credits.
• PopNow 400 naming a field - eleven references, or a ratio not on the list.
• Cancelling stops the wait, not the job - it finishes server-side and bills on success.

Compare it to the local route honestly: BiRefNet in core gets you an excellent matte on most subjects and dies on hair, glass and motion blur - the same failure the whole field has. A model that generates the composite with alpha attached skips the matte problem by never having one. That's worth 4 credits when the job is a product shot on white.