seedream-v4.5

Seedream is ByteDance's closed image line - you can't download it, full stop. The open-source route to it doesn't exist, which is exactly why this node exists: it wraps the doubao-seedream-4.5 API so the model sits in your ComfyUI graph like any generator, with 4K output and multi-image input that your local SDXL checkpoint can't dream of. If you've ever wanted Seedream's strong Chinese text rendering and subject-consistent multi-image fusion without leaving the canvas, this is the door.

How it works

Standard pack wrapper: your prompt (and optionally your images) get base64'd and POSTed to the single gateway endpoint with the model id doubao-seedream-4.5. The server runs the generation and hands back one or more image URLs, which the node downloads and converts to IMAGE tensors. Two things are worth knowing before you run it.

First, multi-image fusion: image_input accepts a whole batch of images, and unlike most packs there's no manual stitching step - you just feed a batch (up to 10 images) and Seedream composes them itself. That's the "multi-photo combination" the author pushes, and it's genuinely useful for e-commerce hero shots where you want product + context from separate photos.

Second, the max_SetImage toggle ("off"/"auto"). Off means one image out. Auto switches the backend into set-generation mode, which produces a grid of related images (the code asks for up to 15). It's not a "number of images" dial in the obvious sense - think of it as "give me a whole variation set."

The inputs that matter

• prompt (STRING) - plain language, and Seedream handles Chinese natively, which is a real differentiator over the US image models.
• custom_size + size + width/height - with custom_size off (default), you pick from 16 preset aspect/quality combos in size (2K:2048x2048, 4k:4096x4096, portrait/landscape variants, and so on). Flip custom_size on and width/height take over - but note the minimum is 1024 on both axes, so you can't go smaller than 1K.
• max_SetImage - the set-generation switch above.
• seed (INT, default 0) - reproducibility.

Optional: image_input (IMAGE batch). One output, output (IMAGE).

Installing it

It's part of comfyui-MJAPI-party. ComfyUI Manager → search "mojieapi_party" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MoJIeAIGC/comfyui-MJAPI-party.git

Then register at mojieaigc.com and add your key to config.ini (leave BASE_URL = https://www.mojieaigc.com/v1/completions alone):

[API]
KEY = your_api_key
BASE_URL = https://www.mojieaigc.com/v1/completions

Restart. No models to download; the deps are the standard numpy/torch/pillow/requests set.

Common issues

• 4K jobs are slow and cost real money. The code's timeout is 20 minutes for a reason. Check your account balance before queueing a batch of 4K renders - the "pro" flag is hardcoded on, so you're on the premium pricing tier whether you noticed or not.
• Error image with text. Read it - usually a bad key, empty balance, or an input resolution the API rejected.
• Images leave your machine, moderation follows ByteDance. Seedream refuses what Seedream refuses, and no node setting bypasses it. That's the standing trade of the whole closed-model-via-API approach (the reseller/proxy layer is a crowded market for exactly this reason).