Seedream5.0 Flash - Text to Image

There is no Seedream you can download. ByteDance's product line - Seedream for images, Seedance for video - is API-only, kept closed while the same company ships Depth Anything, Hyper-SD and Bernini for free. If you want the model, you call it. This is the calling.

The node comes from a family ComfyUI-KIE-Nodes-Next generates out of KIE.ai's API catalog: one ComfyUI node per KIE model, that model's documented request fields as real widgets, and native Comfy media sockets instead of URL plumbing. Prompt in, one image out, no local VRAM.

Why route through KIE.ai at all

ComfyUI has official Partner Nodes for Seedream. People still build packs like this one, and the recurring reason is price. The thread that sticks in memory is a dev who built a KIE.ai node for a competing ByteDance image model purely because the reseller rate beat the first-party one - roughly 40% cheaper at 1K/2K, 60% at 4K, in their own comparison. Same model behind the API; different meter.

The honest trade-off is the one every API node carries: it holds your credential and calls the network by design, which is the shape of the thing that got weaponised in the ComfyUI_LLMVISION incident. So read any API pack before installing it. This one keeps the key out of your workflow JSON - it lives in ComfyUI's user config area, is validated against KIE before it overwrites a working key, and the frontend only ever sees a masked fingerprint.

Mechanically it's an HTTP client with a nice costume: it builds the request from that model's documented schema, submits a task to KIE, polls until it finishes, then downloads the result URL and decodes it into a ComfyUI image tensor. No local sampling happens at any point.

Inputs that matter

• prompt - 3 to 5,000 characters. Under three and it's rejected; the character ceiling is generous enough for dense art-direction prompts.
• aspect_ratio - eight presets: 1:1, 4:3, 3:4, 16:9, 9:16, 2:3, 3:2, 21:9. Yes, 21:9 is in there; no, there's no free-form width/height.
• quality - basic gives 1K, high gives 2K. On a metered API this is a straight cost multiplier, so draft at basic and only re-run at high once the composition is locked.

Optional settings: output_format (png or jpeg), timeout_seconds (default 1200 - your "how long do I wait for the queue" dial), callback_url (leave empty; the node polls for you), and nsfw_checker. That last tooltip is genuinely confusing, so read the widget name literally: turn it on if you want the provider's filter applied. It also warns there's no guarantee either way, which is worth believing.

Outputs

Seven of them. image is the downloaded result as a ComfyUI IMAGE - the one you wire onward. url is the KIE-hosted result, all_urls_json carries every URL the response returned, task_id is the handle you'll want in a minute, and raw_json is the full provider response for when something looks wrong. credits_consumed and credits_left are floats, and the node prints spend and remaining balance on itself after a run.

Keep that task_id. If your graph errors after the job succeeded, do not re-run and pay twice - feed the ID into KIE • Reuse Completed Image, which retrieves a finished task without resubmitting it.

Installing the pack

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
cd ComfyUI-KIE-Nodes-Next
pip install -r requirements.txt

Restart ComfyUI. Requirements are requests, Pillow, numpy, PyYAML - no torch extensions, no model downloads, no OpenCV for this node.

Then set your key once: Settings → KIE.ai Nodes Next → Connection, paste it, and wait for the CONNECTED confirmation. Environment variables are deliberately unsupported; the key lives in ComfyUI's user config directory.

Gotchas

The node isn't in your menu. These catalog-driven nodes are built from KIE's live API docs, so a fresh install needs a sync. Restart ComfyUI and let the catalog fetch run; if it's still absent, add the KIE Next · Catalog Sync node from the Advanced folder and run it. Don't search for the class name - the trailing hex in seedream_5_flash_text_to_image_efb8941f is a stable hash of the docs URL, not something to type. Browse KIE Next / Image / ByteDance / Seedream instead.

A run costs money every time you queue it. There's no dry-run. Right-click any KIE node and pick KIE Next · Check inputs & model controls to inspect documented fields and limits without uploading media or submitting a job - a free preflight before the paid one.

Long waits then a timeout. Raise timeout_seconds. The client already retries transient 429/5xx responses with backoff, so a timeout usually means genuine queue time, not a flaky request.

Vague refusals instead of an image. That's provider moderation, not the node. No local workaround exists - there are no weights to strip.