Nodes/PopNow/Seedream 5.0 Pro · Edit
ComfyUI Node

Seedream 5.0 Pro · Edit

Ten references, transparent output, and a thinking pass

By popnow-ai·Created 24 days ago·Updated about 20 hours ago· 0
Seedream 5.0 Pro · Edit
  • image_urls
  • images
  • alpha
  • task id
◄prompt►
◄aspect_ratioauto►
◄resolution2k►
◄output_formatjpeg►
◄backgroundopaque►
◄enable_thinkingtrue►

Pro Edit is the top of ByteDance's Seedream 5.0 line, and the thing that distinguishes it from the Flash and Lite versions isn't the reference count - it's enable_thinking, a switch that runs a prompt-refinement pass before the image is generated. On by default. It's why this node can take a lazy one-line prompt and still come back with something coherent, and it's also why it takes longer.

Around 5.4 credits (~$0.054) per image, so roughly 20% more than Flash. If you're editing one important frame rather than iterating on fifty, that's not a meaningful difference.

What it does

PopNow's node for bytedance/seedream-5.0-pro/edit. image_urls is an autogrow socket group - one connection per reference, up to the model's limit of ten. Your sources upload on queue and are cached by content hash for 12 hours, so a prompt-tweaking loop doesn't re-upload the same images. The result comes back as a ComfyUI IMAGE batch, with an alpha mask batch alongside because this model's schema has a background property.

Closed model, hosted call, metered per image, refunded on failure. Same terms as everything else in this pack.

The inputs that matter

  • prompt - the image you want back: subject, composition, style, lighting. With enable_thinking on, you can be terser than you would be elsewhere; the pass fills gaps. Turn it off and you'll want to write like you mean it.
  • image_urls - up to ten sources. Prompt in slots ("person from image 1, the fabric from image 3") - multi-reference edits get vague fast.
  • aspect_ratio - auto (default; picks the supported ratio closest to your first reference) or an explicit one from 1:1 up through 21:9, including 4:5, 5:4, 2:1 and 1:3. auto is usually right for an edit.
  • resolution - 1k or 2k. Two tiers, not three. Note this is a per-image price that doesn't change with the tier.
  • output_format - jpeg or png. png for the transparent path.
  • background - opaque or transparent. The tooltip spells out the requirement: png output, and one PNG input with alpha.
  • enable_thinking - on by default. Turn it off when you know exactly what you want and the extra pass is just latency.

Outputs, and one trap

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

The RGB the pack hands you is premultiplied against alpha for transparent results, which is what keeps a cutout looking right in previews. So don't multiply by the mask again - if you wire alpha into a compositing node and the image still has its premultiplication, you get dark fringes. Take images as-is; use alpha only where a node genuinely wants a mask.

If alpha comes back all ones, one of the two transparency requirements is missing: background isn't transparent, or output_format isn't png.

Install

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

Or install PopNow from ComfyUI Manager. Key from popnow.ai/api/console/keys → Settings > PopNow (or POPNOW_API_KEY), then restart again so the catalog populates the node list. No pip dependencies: the pack ships an empty dependency list and uses only what ComfyUI already includes.

Common issues

  • PopNow 40201 - out of credits; at ~5.4 a frame, a batch of thirty is 162.
  • PopNow 40004 - unsupported combination. With this node, the transparent background plus a non-PNG source is the classic.
  • PopNow 400 naming a field - eleven references, or a ratio not in the list.
  • content_rejected / generation_failed - model-side decline or failure, refunded. Reword rather than resubmitting identically.
  • An error ending in request_id ... - exact identifier for support; skip the guessing.
  • Cancelling stops the wait, not the job - it runs to completion and bills on success. Use Settings > PopNow > Max credits per run as the guard that actually works.

Compared with doing this locally, the honest trade is: no Qwen-Image-Edit graph, no ControlNet stack, no LoRA training, and no chance of reproducing a character exactly next month. Pro Edit is very good at this frame from these references. If you need the same face across an entire project, that's still a trained LoRA's job - this is what you use when the project is due Friday.

CategoryPopNow/Image/ByteDance

Inputs (7)

NameTypeDefaultDescription
promptSTRINGDescribe the image to produce from the source: subject, composition, style, and lighting.
image_urlsCOMFY_AUTOGROW_V3Source images that you are authorised to process.
aspect_ratioCOMBOautoOutput aspect ratio. auto picks the supported ratio closest to the first reference image.
resolutionCOMBO2kOutput resolution tier.
output_formatCOMBOjpegFile format of the returned image.
backgroundCOMBOopaqueOpaque image, or a PNG with a transparent background (png output, one PNG input with alpha).
enable_thinkingBOOLEANtrueRun a thinking pass that refines the prompt before generating. Off is faster.

Outputs (3)

NameTypeDescription
imagesIMAGE—
alphaMASKTransparency of each image: 1 is opaque.
task idSTRINGThe PopNow task id, for the API console logs or support.