RunPod WAN 2.1 T2V 720p

Text-to-video is the mode where Wan 2.1 goes from a prompt to a clip with no input image at all, and this node is how you do it without downloading a 14B checkpoint or clearing half your SSD. You paste an API key, write a scene, and a hosted wan-2-1-t2v-720 endpoint renders it. If your hardware can't hold a video model - or you just don't want it to - this is the whole workflow in one node.

The catch you should hear up front: T2V is the harder mode. With image-to-video you anchor the first frame and the model has something concrete to hold onto. With T2V, composition, motion, and consistency all come out of the prompt alone, so results are less predictable and subject drift is real. The community's default is I2V for exactly this reason. This node is for when you don't have a starting image, or when pure prompt-to-scene is the whole point.

How it works

Same machinery as every node in this pack. The node is built by a factory function, so its behavior is identical to its I2V sibling minus the image handling: it submits an async job to https://api.runpod.ai/v2/wan-2-1-t2v-720/run with your key in the Authorization header, then polls the status endpoint once a second. The console shows [RunPod] Status: IN_PROGRESS lines as it waits. Nothing runs locally - no weights, no VRAM, no VAE to keep straight.

Wan 2.1 is flow-matching, so the sampler knobs are guidance and flow_shift rather than SDXL-style CFG. The defaults of 5 and 5 sit in the right zone (the community sweet spot for Wan guidance is roughly 5–7).

The inputs that matter

The required list is a trimmed version of the I2V node - there's no image input anywhere, which is exactly the point:

• api_key - your RunPod key, straight into the field.
• prompt - the whole scene in a sentence or two. The default is a good example of what works: "Steady rain falls on a bustling Tokyo street at night, neon signs casting vibrant pink and blue light" - subject, motion, lighting, mood.
• duration - 1–10 seconds, default 5.
• seed - -1 for random; set it to lock a take.

negative_prompt, size (1280*720), num_inference_steps (30), guidance (5), flow_shift (5), enable_prompt_optimization, and enable_safety_checker fill out the rest. The safety checker is on by default and lives server-side.

The output

One output, video_url, a STRING. Wire it into a "Load Video from URL" node or Save Text it. The URL expires after roughly seven days - download anything you want to keep, and don't be surprised if a saved link you revisit later is dead.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

Restart ComfyUI (or install "RunpodNodes" via ComfyUI Manager). Zero extra Python dependencies - the pack uses only what ComfyUI already ships - and no model downloads, because the model never touches your disk.

Worth remembering before you make this a habit: every render is a paid API call (the README's video range is about $0.30–1.20), your prompt leaves your machine, and the pack is a brand-new single-author release, so give the code a skim before pasting in a key. For pure prompt-to-scene on a machine with no GPU headroom, though, it's hard to beat.