ByteDance Seedream 4.5 & 5.0

ByteDance is the strangest lab in this space: its flagship generation models are closed without exception - Seedream and Seedance are API products, full stop - while its components ship open and quietly become load-bearing infrastructure. The KB's ByteDance panel sums up the split: anything it charges for stays on the API.

The Seedream node is the sanctioned way to use one of those API-only flagships inside ComfyUI. It's the full Seedream line - 4.0, 4.5, 5.0 Lite and 5.0 Pro - behind one dynamic picker, up to 4K output, and it does something rare for a text-to-image node: single-sentence editing of existing images, from reference images you feed it.

What it actually is

A core partner node - ships with ComfyUI, no install, cloud-backed through Comfy's proxy to ByteDance's Seedream API. Sign in with your ComfyUI account, spend credits, and the generation runs on ByteDance's servers. Synchronous submit-and-poll: you get a normal IMAGE tensor back into your graph. No weights on your disk, no 24GB VRAM requirement.

The inputs that matter

• prompt - for creating an image or, with references attached, editing one. Seedream's editing is famous for following a single precise instruction - "change the shirt to a red one, keep everything else" - without the mask ceremony.
• model - the DynamicCombo: seedream 5.0 pro, seedream 5.0 lite, seedream-4-5-251128, seedream-4-0-250828. Newer models cost more and hold to stricter resolution floors. Each variant opens its own sub-options:
  • size_preset - recommended sizes, or Custom to use width/height.
  • width / height - custom dimensions (with per-model minimums and maximums; 5.0 Lite and 4.5 need at least ~3.7MP, and 5.0 Pro caps around 4MP, while the lite line goes up to 16.8MP - actual 4K).
  • max_images - generate up to N related images in one run (story scenes, character variations). Total reference + generated images can't exceed 15.
  • images - up to 10 reference images (14 on 5.0 Lite). Feed one in and the prompt becomes an edit instruction; feed several for multi-reference generation.
• seed - reproducibility within the same model and settings.
• watermark - off by default; flip it on to stamp the image with an "AI generated" watermark.

Where it fits

Seedream's reputation is two things: exceptional text rendering (it handles in-image text better than most of its generation) and the editing capability. If you've ever been stuck rerolling a whole image because one sentence of text came out garbled - Seedream 4.5/5.0's reference-based editing lets you fix the text in place. The 4K ceiling also makes it a one-shot base layer: generate at full res in the cloud, then skip the local hi-res fix and go straight to fine detail work.

Gotchas

• Resolution floors are real and enforced: pick a size preset under the model's minimum and the node errors out before you spend anything, but it's a common first-run stumble. Use the presets, don't freehand width/height until you've read them.
• Reference images must keep an aspect ratio between 1:3 and 3:1.
• Per-call pricing differs by model and resolution, and the node shows an estimate before you queue - 5.0 Pro is the premium rung for a reason.
• The partner-node standard caveats apply: Comfy account + credits, and login is localhost-only, so a LAN --listen setup will fight you.

If you want ByteDance's best without waiting for them to open the weights - which, per their own track record, they may never do - this is the door. Text rendering, 4K headroom, and single-sentence editing are the reasons to walk through it.