Wan Image-to-Video Generator

Image-to-video is the mode the Wan community actually runs. Anchoring the first frame to a known image solves the consistency problems that plague pure text-to-video - you already know what the scene looks like, so the model just has to move it. WanI2VGenerator ("Wan Image-to-Video Generator") from ru4ls/ComfyUI_Wan is that workflow on Alibaba's DashScope API: give it a picture URL and a motion prompt, get back a ~5-second mp4.

The whole point of this pack is that you don't need a video-generation GPU to do this. Local Wan 2.2 I2V on a 4090 runs 15-40 minutes a clip at 720p; this node hands it to Alibaba's servers and bills you per job. Trade the VRAM anxiety for a credit card and a polling loop.

How it works

Classic async DashScope job: submit image_url plus your prompt to the video-synthesis endpoint with X-DashScope-Async: enable, get a task_id, poll every 10 seconds (up to 60 attempts - video takes a while), then download the finished mp4 to your local output_dir. Outputs are video_file_path and video_url; to actually see the clip in ComfyUI, feed the path into a "Load Video (Path)" node from ComfyUI-VideoHelperSuite. The README says this for every video node, and it's the step everyone forgets.

The inputs that matter

• model - three hosted variants: wan2.5-i2v-preview, wan2.2-i2v-flash (default), and wan2.2-i2v-plus. Flash for speed, plus for quality, preview for the newest. If you're iterating on a prompt, flash is your friend; for the final render, switch to plus.
• image_url - required, and it must be publicly accessible. Alibaba's servers fetch it, so a local file path will silently fail. Host your frame somewhere reachable first.
• prompt - write it like a motion director, not a caption. Wan I2V prompting wants "the character stands up and walks toward the camera," not "a beautiful woman." The KB's Wan essay calls this the George R. R. Martin style - describe the action, not the scene.
• negative_prompt - optional; handy for banning artifacts.
• resolution - 480P, 720P (default), 1080P. 1080P is where cloud cost and wait time climb.
• prompt_extend - on by default; Alibaba's prompt rewriter. Fine to leave on, but if your motion wording is getting mangled, switch it off.
• seed, watermark, output_dir - the usual. Seed 0 is random.

Install and setup

ComfyUI Manager, search ComfyUI_Wan, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ru4ls/ComfyUI_Wan.git
pip install -r ComfyUI_Wan/requirements.txt

Restart, then copy config/.env.template to config/.env in the pack folder and set DASHSCOPE_API_KEY (plus DASHSCOPE_API_KEY_CHINA if you'll use the mainland China region). The dependencies are light - requests, Pillow, numpy, python-dotenv - because generation is remote. It's paid per job; the README flags it in bold.

Gotchas

• It's a 5-second ceiling, same as local Wan. If you need longer, that's what the pack's VACE Video Extension node is for - chain clips or extend.
• First-frame chaining works: take the last frame of one clip as image_url for the next. That's the community's standard long-video trick, and it works here because the I2V anchor is doing the consistency work.
• If a job errors with a task failure, re-check that your image URL is actually fetchable - a broken link up front is the most common reason a run dies mid-poll.

If you've got a 3090+ and want the free, open-weight version of this, local Wan 2.2 is still great. This node is for everyone who'd rather pay a few cents than babysit a 40-minute render.