Bytedance Dreamina Text-to-Image

This is the paid-API route to ByteDance's Dreamina model, reached through WaveSpeed AI's billing. You give it a prompt, it returns a real IMAGE tensor you can keep working with in ComfyUI - no "paste the URL into a browser" step. Dreamina's reputation is for strong text rendering and polished, commercial-friendly output, and the V3.1 variant you get by default is the good one.

Here's the mental model: this pack (ERPK Collection) doesn't run the model locally - nothing downloads, no VRAM pressure. The node talks to the WaveSpeed API, which runs Dreamina, polls the job, converts the returned image to a tensor, and hands it to the next node. It's exactly the "cloud service in a ComfyUI node" pattern, and it works because WaveSpeed is a reseller of ByteDance's models, so you get one API key for a lot of different generation services.

Inputs that matter

• model - Dreamina V3.1 (default) or Dreamina V3.0. V3.1 is the newer variant; switch down only if you have a specific reason.
• prompt - the actual image description. Multiline.
• width / height - both default to 1328, range 512–2048 in steps of 8. Match your downstream latent size to avoid a resize surprise later.
• seed - -1 randomizes for a fresh billed generation each queue; a fixed value is cache control, letting ComfyUI reuse the previous result.
• enable_prompt_expansion - on by default. WaveSpeed rewrites/expands your prompt for better results. It genuinely improves output, but it means the model isn't seeing your literal words - turn it off if you're trying to nail something precise.
• enable_sync_mode - off by default. When off, the node returns immediately with a job reference and ComfyUI waits; on, it blocks until the image is done. You almost never need to touch this.
• enable_base64_output - off. Leave it; the node converts to a tensor either way.

The single output, image, is a standard IMAGE tensor - wire it to Preview Anything, a Save node, an upscaler, whatever.

Install and billing reality

ComfyUI Manager → search erpk → ERPK Custom Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Restart, then add a WaveSpeed API key (wavespeed.ai) under Settings > ERPK > API Keys. Every generation is a paid API call, billed per image by WaveSpeed - there's no free tier to fall back on, so the seed trick (fixed seed = cached, un-billed rerun) is worth remembering.

People get burned by two things. First, forgetting enable_prompt_expansion is on and wondering why the output doesn't match the prompt. Second, treating the 1328 default as sacred - it's not a model-native size, so if your downstream expects 1024, set it explicitly and save a resize.