Bytedance Seedream V4.5
The text-and-typography specialist, up to 4K
- client
- image
Seedream V4.5 is ByteDance's image model built for the one job most diffusion models are terrible at: rendering legible text. Posters, logos, UI mockups, menus, title cards - if the image has to contain words that actually spell something, this is the model to reach for. And unlike the Qwen nodes in this pack, it goes all the way up to 4096px, so you can produce print-sized output, not just screen-sized.
It runs through the WaveSpeed AI API via the ERPK Collection, the community pack that wraps the service for ComfyUI. Cloud-billed per image, no local weights.
How it works
Type a prompt, pick a size, and the node sends the job to WaveSpeed's Seedream V4.5 endpoint, then downloads the result into a ComfyUI image IMAGE tensor. The resolution story is the headline: 1024 to 4096px, with one-click presets for the common ratios - 1:1 at 2048, a 4K 1:1 at 4096, plus 4:3, 3:4, 3:2, 2:3, 16:9, and 9:16. The size_preset combo is the clean way to use it; switch to Custom and the width/height fields kick in.
The inputs that matter
prompt- the image to generate. V4.5 rewards prompts that specify the text content in quotes and its placement: a coffee shop menu board that reads "OPEN", bold serif, centered - it's a text-rendering model, so spell out the words.size_preset- resolution presets, Custom default. The presets are the fastest path to a specific ratio; Custom is for exact dimensions.width/height- 1024–4096, step 8, default 2048. Only used whensize_presetis Custom.seed- cache control, randomizes by default. A fixed seed reuses a cached result and skips re-billing; randomize forces a fresh generation. This isn't a real seed - it's the ERPK cache knob.show_aspect_ratio- cosmetic, shows the ratio in the node title.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt
Or ComfyUI Manager → search erpk → ERPK Custom Nodes. WaveSpeed key with credits via right-click canvas > ERPK Settings (Settings > ERPK > API Keys).
Common issues
- Garbled text in the image - V4.5 is good, not psychic. Short words, high contrast, and a quoted exact string in the prompt go a long way; long paragraphs will still glitch.
- "Prompt is required" - free error, node validates before billing.
- Fixed seed returning the same image - cache control, by design. Randomize to actually generate.
- 4K is slow - 4096px renders take a while; the node polls until done. Budget the wait for print-size output.
- Cost - 4K-class renders cost more than the small Qwen tier. Use the resolution you need, not the maximum you can afford once.
This is the pack's typography workhorse. If your project needs a poster, a logo, or anything with words, V4.5 is where the money goes; for plain illustration, the Qwen nodes are cheaper per unit.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| prompt | STRING | Text description of the image to generate | |
| size_preset | COMBO | Custom | Recommended resolution presets. Select 'Custom' to use manual width/height. |
| seed | INT | -1-1–2147483647 | Seed for cache control. Randomizes by default. |
| clientopt | WAVESPEED_AI_API_CLIENT | WaveSpeed API client (optional if API key is configured in Settings) | |
| widthopt | INT | 20481024–4096 | Custom width (only used when size_preset is 'Custom') |
| heightopt | INT | 20481024–4096 | Custom height (only used when size_preset is 'Custom') |
| show_aspect_ratioopt | BOOLEAN | true | Show aspect ratio in node title |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |