RunPod Seedream 4.0 T2I

Seedream 4.0 is ByteDance's current-generation text-to-image model, and like its sibling Seedream 3.0 it's closed - no weights, no local install, no fallback. This node is the seedream-v4-t2i endpoint wrapped as a ComfyUI node, so the newest Seedream is one node and one API key away. If you're reading this because you want the 4.0 look without a ByteDance account and a separate SDK, this is the path.

It's a "new-generation text-to-image" per the pack's own framing, which mostly means better prompt adherence and a sharper default aesthetic than 3.0. Whether that's worth per-image money over an open local model is your call - the honest take is that Seedream 4.0 is genuinely good at following complex prompts, and closed means the only way in is through something like this node.

How it works

Standard pack mechanics: async POST to seedream-v4-t2i, poll until done, base64 image back out to an IMAGE tensor. Your GPU does none of it. The console logging ([RunPod] Status: ...) is your only progress indicator, so keep an eye on it.

The inputs that matter

• api_key - RunPod key.
• prompt - the generation text; this is where 4.0's adherence shows. Complex, specific prompts are the point.
• negative_prompt - filled in and useful; Seedream has real CFG so it actually does something.
• size - string, default 1024*1024. Keep the * separator format.
• seed - -1 random; set it to reproduce a keeper.
• enable_safety_checker - on by default, enforced server-side.

What comes out

One image output (IMAGE). Save it, preview it, or feed it onward - standard tensor behavior.

Installing it

One of the 27 nodes in tcarwash/ComfyUI_RunpodNodes. ComfyUI Manager: search ComfyUI_RunpodNodes, or

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

then restart. No requirements.txt to satisfy - the pack's only real deps (requests, Pillow, torch) are already in ComfyUI. Get a RunPod API key at runpod.io.

Gotchas

• Closed + metered - the double whammy. You're renting a model you can never own, and the meter runs every render.
• Queue - 30–120 seconds of waiting is normal for public endpoints; it's not hung.
• "Job failed" - key or credits, in that order.
• Filtering at the source - safety checker on, no bypass possible, no weights to patch.

If you want the editing side of Seedream 4.0, skip this node and grab RunPod_Seedream4Edit in the same pack - it takes up to two input images plus an instruction. But for straight generation with ByteDance's latest prompt-following, this is the one.