Nodes/Seedream4.5-comfyui-doubao4.5-comfyui/🚀 Seedream 4.5 Image2Image
ComfyUI Node

🚀 Seedream 4.5 Image2Image

The Image2Image API Node

By ninepoin4·Created 2 months ago·Updated 2 months ago· 0
🚀 Seedream 4.5 Image2Image
  • image
  • image2
  • image3
  • image4
  • IMAGE
prompt保持模特姿势不变,改变材质...
api_key
modeldoubao-seedream-4-5-251128
size2K
watermarkfalse
timeout30
sequential_image_generationdisabled
max_images1

The "🚀 Seedream 4.5 Image2Image" node is a wrapper that calls ByteDance's Seedream 4.5 through the Volcengine Ark API - so it runs nothing locally. No checkpoint, no VAE, no VRAM, no CUDA. You feed it an image and a prompt, it phones ByteDance's cloud, and a generated image comes back as a normal IMAGE tensor you can wire into the rest of your graph. For the Image2Image/editing use case this is worth taking seriously: when Seedream 4.5 shipped in late 2025 it ranked #3 on the image-edit leaderboard, and it was in high demand in ComfyUI precisely because it has no open weights. This pack filled that gap two weeks later, in December 2025.

What it actually does

Like every API-wrapper node, this one is an HTTP client dressed as a generator. Under the hood it takes your ComfyUI IMAGE tensors, encodes each as a base64 PNG data URL, and POSTs them - plus your prompt and settings - to https://ark.cn-beijing.volces.com/api/v3/images/generations with your key as a Bearer token. The API replies with URLs, the node downloads the images, converts them back to float tensors, and hands you one IMAGE output. From the canvas it looks like a local sampler did the work; underneath, your prompt and reference images just left your machine.

The default model is doubao-seedream-4-5-251128. That model name, the endpoint, and the whole workflow are things you don't touch - this isn't a "hack the config" node, it's a paste-a-key-and-go one.

The inputs that matter

  • image (required) - your reference image. Seedream is genuinely good at "keep the subject, change the scene/material," which is what the default prompt (保持模特姿势不变,改变材质..., "keep the pose, change the material") points at.
  • image2/image3/image4 (optional) - up to four reference images total. All of them go to the API as a multi-image reference array, so you can do things like identity-plus-outfit-plus-background.
  • api_key - from the Volcengine Ark console (console.volcengine.com). It's a password-type field, so it won't be visible on the canvas.
  • size - 1K, 2K, 720x1280, 1280x720; defaults to 2K. The API honors whatever you pick.
  • timeout - 1–300s, default 30. The README is loud about this: set it to something in the 30–100s range before your first run. Generations can genuinely take longer than 30 seconds.
  • sequential_image_generation (disabled/auto/enabled) + max_images (1–8) - only relevant if you want multiple variations out of one call.

Output is a single IMAGE tensor - a batch if you asked for multiple images. Wire it to a Preview/Save node and you're done.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/ninepoin4/Seedream4.5-comfyui-doubao4.5-comfyui
pip install -r Seedream4.5-comfyui-doubao4.5-comfyui/requirements.txt

Or search the pack title in ComfyUI Manager. The dependency list is refreshingly tiny - just requests and Pillow - and there are no model files to download and no GPU needed, because there's no model. Restart ComfyUI and the node appears under the Seedream category. One caveat: the README is entirely in Chinese. It's short and this article covers everything in it, but if you're going to poke at settings yourself, have a translator handy.

Where people get burned

  • The timeout. The README's first line is literally a warning to set it before doing anything else. Start at 60.
  • Silent failure. When the API call fails, the node doesn't crash the workflow - it catches the exception, prints to the console, and returns the input image as if nothing happened. If your output suddenly looks identical to your input, that's not a creative decision; that's the error path. Watch the ComfyUI console.
  • Cost and keys. This is metered, per-call inference. You need credit on your Volcengine account or every request fails. Seedream 4.0 was known as cheap, but "cheap per call" adds up fast if you batch.
  • The filter. Seedream is a ByteDance-hosted model, and you can't abliterate a model you never downloaded. Whatever ByteDance moderates, this node can't get around - so don't route anything through it you know the filter won't pass.
  • The security posture. A node whose job is to hold your API key and phone the network is exactly the shape of thing this ecosystem got burned by once. The code is three files and requests+Pillow, so glance at it before you paste a paid key in. And note it's not the official ComfyUI Partner Node path - it's a direct-to-Volcengine wrapper, which is the point if you want your own key and billing.

Use it when you need a closed model inside your workflow - because Seedream 4.5 is only ever going to be an API, and this is the cheapest door in.

CategorySeedream

Inputs (12)

NameTypeDefaultDescription
imageIMAGE
promptSTRING保持模特姿势不变,改变材质...
api_keySTRING
modelSTRINGdoubao-seedream-4-5-251128
sizeCOMBO2K4 options: 1K, 2K, 720x1280, 1280x720
watermarkBOOLEANfalse
timeoutINT301–300
sequential_image_generationCOMBOdisabled3 options: disabled, auto, enabled
max_imagesINT11–8
image2optIMAGE
image3optIMAGE
image4optIMAGE

Outputs (1)

NameTypeDescription
IMAGEIMAGE