Wan Text to Image

Wan is ByteDance's model family - best known in this community for open video weights you can run locally. This node is the other side of that company: the hosted text-to-image API, model wan2.5-t2i-preview, served through ComfyUI's partner-node system. Prompt in, tensor out, no VRAM spent.

The ByteDance pattern is worth knowing before you use it: components open, products closed. You can run Wan 2.1/2.2 video on your own GPU, but the flagship image API stays behind a hosted service, and this node is the polite front door to that service. What you get is a capable, fast image model with two genuinely useful features you don't see on every hosted node: prompt_extend, an AI prompt-enhancer that rewrites your prompt into something stronger (on by default), and bilingual prompting - the prompt field officially supports English and Chinese, which is a reminder of where this model comes from and a genuinely handy feature if you work with Chinese-language prompts.

The inputs that matter

• model - combo, currently just wan2.5-t2i-preview
• prompt - multiline, English or Chinese
• optional negative_prompt - what to avoid
• optional width / height - 768 to 1440, step of 32, default 1024×1024
• optional seed - actual seed for generation
• optional prompt_extend - boolean, default on. "Whether to enhance the prompt with AI assistance." Leave it on and your short prompts get fleshed out; turn it off if you want literal control.
• optional watermark - boolean, default off. Adds an AI-generated watermark to the result if you want the provenance marker.

Output is a standard IMAGE tensor.

How it works

Your request is submitted to /proxy/wan/api/v1/services/aigc/text2image/image-synthesis, which returns a task ID; the node then polls the task endpoint every few seconds until the image is ready and downloads it. Task-based, like Runway - expect queue latency, not instant returns. Billed per call from your Comfy account credits; ships with ComfyUI core, no install, no model files.

Common issues

• Prompt extend can fight you. It's great when your prompt is thin; it's annoying when you've written something specific and the enhancer "improves" it into something different. If outputs drift from intent, flip prompt_extend off and write longer prompts yourself.
• It's a hosted preview. "Preview" in the model name is doing real work - this is ByteDance's hosted pipeline, versioned server-side, and you have no local fallback through this node. If the service changes, the node changes.
• Aspect control is coarse. Width/height step of 32 between 768 and 1440 is flexible enough, but there's no one-click ratio preset - set both explicitly for the shape you want.
• Slow-ish by design. The polling means a run holds the graph open for a bit. Not broken; hosted-queue reality.

Reach for it when you want a ByteDance-tier image model without the ByteDance web app, when you're working with Chinese-language prompts, or when you want the prompt-extend convenience of a hosted model inside your ComfyUI graph. It's a clean, honest hosted t2i node - just remember the "preview" and the queue.