🎬 Wan Video Generator

Straight to the point: Wan Video Generator does not run Wan. The pack's source loads a DiffusionPipeline from damo-vilab/text-to-video-ms-1.7b, an old 2023 ModelScope text-to-video model, with a code comment that reads "Replace with actual Wan model when available." The comment is still there. When people say "Wan" in the local video world in 2026, they mean Wan 2.1/2.2 - a 14B-class model with a serious LoRA and ControlNet ecosystem. This node is a placeholder wearing a Wan costume.

That doesn't make it useless. It makes it useful-with-correct-expectations. If you want a quick test clip to prove a pipeline works - wire up the frames, the format, the save step - without waiting for a 14B model, this will do it in a couple of minutes. If you expect Alibaba-grade motion quality, you'll be disappointed fast.

What you set

• prompt and negative_prompt - text controls, same as any T2V.
• num_frames - 4 to 64 (step 4). 16 is the default and a reasonable test length.
• width/height - 256 to 1024, step 64. 512×512 default.
• fps - 1 to 30. Note this is metadata for how you save; the model itself just produces frames.
• guidance_scale and num_inference_steps - 7.5 and 50 are the defaults; 1.7B-scale models don't need heroic step counts.
• seed - -1 for random, otherwise lock it for reproducible takes.

Output is a batch of IMAGE frames (the schema's output) - you stack, save, or feed them into a video output node yourself.

The honest mechanics

It runs through diffusers, fp16, on CUDA if available. First run downloads the model (a real chunk of GB), then it's cached. And here's the pack's signature behavior: if the model load or generation throws, the node doesn't error - it returns placeholder frames, a gradient that shifts color per frame. You can generate a "video" that is literally a color-cycling rectangle and never see a stack trace. Always check the console output for ✅ Generated N frames versus an ❌ Error line.

Installing it

Same pack as everything else - 24oiduts-ComfyUI:

• ComfyUI Manager: search 24oiduts, install, restart.
• Manual:

  cd ComfyUI/custom_nodes/
  git clone https://github.com/GeekyGhost/24oiduts-ComfyUI
  pip install -r 24oiduts-ComfyUI/requirements.txt

  then restart.

diffusers is the dependency this node actually needs. The GitHub README is a stale Studio42 template that says "not recommended for use at this time" - treat it as unmaintained docs, not a verdict.

If you actually want Wan

This is the important paragraph. Real Wan 2.1/2.2 runs in ComfyUI through its native support and Kijai's ComfyUI-WanVideoWrapper, with real T2V/I2V, LoRAs, and ControlNet. Those models are 14B and want a serious GPU, which is presumably why this pack stubbed the name instead. If your goal is actual Wan-quality video, don't build your workflow around this node - go get the real thing. If your goal is a placeholder that keeps the graph wired while you're still planning, this works.

Troubleshooting

• Color-shifting rectangle output - an exception was swallowed; read the console.
• Model downloads on first run - expected.
• Slow/crashes - the underlying 1.7B model still needs a few GB of VRAM; lower width/height/num_frames.

The node name oversells it, but as a lightweight T2V stub it's honest enough - as long as you know going in that "Wan" is aspirational.