Seedream 官方输入 (6000px)

If you feed ComfyUI's built-in ByteDance Seedream node a 4000px photo as a reference, it hands the API a 2048x2048 version of it and never tells you. That's the entire reason this pack exists. Same API, same billing, same model list - the difference is that reference images go up at BytePlus's official limits instead of Comfy's silent downscale.

What it actually changes

ByteDance is the clearest two-track lab in this space: every Seedream and Seedance product is API-only, while the components underneath (PuLID, SDXL-Lightning, Depth Anything) ship free. So Seedream is a closed model you call, not one you load. Comfy's Partner Nodes are the vetted, metered path to it - you're logged into a Comfy account, spending prepaid credits, and your prompt and reference images leave your machine.

This node is a subclass of the core Seedream node. It doesn't route around the official endpoint or resell anyone's key. It just prepares the upload properly: longest side up to 6000px, total pixels up to 36MP, file under 30MB (it switches the upload to JPEG when a reference is large), and a check against the official 1:16 to 16:1 aspect window.

One thing to get straight before you install it: output size does not change. Seedream 5.0 Pro still returns roughly 2K / 4.19MP. More input pixels buys you detail the model can read - small text, faces, logos, fabric - not a bigger file back. People install this expecting 6000px generations and then complain the README lied. It warned you.

There's also a real bug fix buried in here. The built-in node caps the Pro model's height fields at 2496, which clips the tall end of 9:16 at full resolution (1536x2730). This pack raises that cap to 4096.

The inputs you touch

prompt is your multiline instruction - text-to-image or an edit, and it can't be empty. model is a dynamic combo: pick the Seedream version and it expands the size and reference-image fields underneath, where the actual output dimensions and your uploaded images live. seed goes straight through to the API's seed parameter - the same value, same prompt, same reference should come back the same.

watermark defaults to false and just controls the "AI generated" corner mark - leave it off unless a platform requires it.

thinking is the one worth thinking about. It defaults to true and enables the model's prompt-optimization reasoning pass, which genuinely improves adherence on complicated prompts and genuinely makes you wait - noticeably on Seedream 5.0 Pro. You can only switch it off for pure text-to-image; if you supply reference images, the API requires it on, and the node will raise rather than silently ignore you.

The output is a single IMAGE. It wires into SaveImage, a preview, or whatever local upscaler or compositor you're running next.

Install

ComfyUI Manager → search Seedream Official Input, or search the node itself for Seedream 官方输入 (6000px). Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/NoNameFaces/ComfyUI-SeedreamOfficialInput

Then restart ComfyUI (or Comfy Desktop). There is no requirements.txt in the repo and nothing to pip install - it leans on the environment ComfyUI already ships. No model downloads either, because there are no weights; it's an HTTP client with your account attached.

Where people get burned

Your ComfyUI is probably too old. The node imports the Seedream thinking-optimization options straight out of core's comfy_api_nodes package. The thinking widget only landed in stock ComfyUI around v0.28.0. On an older build you get an import error and the node never appears. Update ComfyUI first, then install.

Credits and account. This is a Partner Node: no key to paste, but no free tier either, and the calls are metered per image. Budget accordingly if you're iterating.

Reference limits. Up to 10 reference images, 14 on the newest 5.0 model id, and once you're generating multiple images per call the references plus generations can't exceed 15. The node raises a clear error rather than truncating.

Privacy. Your inputs go to ByteDance's servers for processing, under their moderation policy. If the whole point of your workflow is that nothing leaves the machine, this node is the wrong tool and no amount of node-swapping fixes that.

If you're using Seedream for high-res references - garment shots, product renders, dense text - this is the version to install. If you're doing plain text-to-image with no reference image, it buys you nothing and you can stay on the built-in.