RunPod Seedream 3.0

Seedream 3.0 is ByteDance's closed image model, and its marquee feature is being genuinely bilingual - strong text rendering in both Chinese and English, plus a photorealistic base that punches above most open checkpoints. There are no open weights, so local isn't an option no matter how much VRAM you own. This node is the "use it anyway" path: seedream-3-0-t2i as a RunPod public endpoint, plugged straight into your graph.

It's the model from the closed frontier, the kind the KB's API-node doc covers: no download door, no filter removal, no community bypass. What you get instead is a per-image call into a server that does the heavy lifting, and an IMAGE tensor back out that slots into the rest of your workflow like a local render.

How it works

The pack's standard async plumbing: POST prompt and settings to RunPod, poll the job status (you'll see IN_QUEUE → IN_PROGRESS in the console), decode the returned base64 into an IMAGE tensor. Nothing runs locally; the endpoint holds the model.

The inputs that matter

This is one of the leaner nodes in the pack - just five inputs:

• api_key - your RunPod key.
• prompt - the generation text. This is where the bilingual skill lives: prompt in English or Chinese and the model handles text-in-image in both.
• seed - -1 random; set it to lock a result.
• guidance - default 2.0, which is the low-and-let-it-breathe end for Seedream. Higher pushes adherence but can get crunchy.
• size - a string, default 1024x1024. Note the x separator here (vs * on some sibling nodes) - keep the endpoint's expected format.

What comes out

One image output (IMAGE). Save, preview, upscale - it's a plain tensor, chain it however you like.

Installing it

One of 27 nodes in tcarwash/ComfyUI_RunpodNodes. Install via ComfyUI Manager (search ComfyUI_RunpodNodes) or:

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

Restart. No extra Python deps - the pack runs on what ComfyUI already ships. You'll need a RunPod API key from runpod.io.

Gotchas

• It's closed and metered - no weights, no local fallback, per-image billing. The cost of the bilingual text rendering is that you're renting it every time.
• Queue waits - public endpoints idle in line; 30–120 seconds is normal. The console's [RunPod] Status: lines are your progress bar.
• "Job failed" - key or credits, always check those two first with this pack.
• Filtering at the source - ByteDance's safety checker (on by default) applies server-side, and since the model is closed there's nothing to patch.

Seedream 3.0 was a mid-pack release - the 4.0 T2I sibling in this same pack is the newer generation if you want the latest ByteDance behavior. But for bilingual text-in-image at a decent photorealistic floor, this node is the direct route, and you don't have to own a ByteDance data center to use it.