Wan2.1 I2V-14B: separate 480P and 720P checkpoints, size as area, and a table typo

Wan2.1 image-to-video comes as two separate 14B checkpoints. Wan2.1-I2V-14B-480P supports 480P and Wan2.1-I2V-14B-720P supports 720P; the card's resolution table marks each as unsupported at the other resolution. The table names the 480P model "Wan2.1-T2V-14B-480P", but the download table, every command and the Gradio --ckpt_dir_480p flag use Wan2.1-I2V-14B-480P, so that entry is a typo and no text-to-video checkpoint is reused. Setup: clone github.com/Wan-Video/Wan2.1, ensure torch >= 2.4.0, pip install -r requirements.txt, then download with huggingface-cli download Wan-AI/Wan2.1-I2V-14B-480P --local-dir ./Wan2.1-I2V-14B-480P (or modelscope download). Run: python generate.py --task i2v-14B --size 832*480 --ckpt_dir ./Wan2.1-I2V-14B-480P --image <input> --prompt "...". For the 720P checkpoint use --size 1280*720. For image-to-video, size is the output area and the aspect ratio follows the input image. On one GPU the card's efficiency tests used --offload_model True for 14B models. Prompt extension is optional: local with --use_prompt_extend --prompt_extend_model Qwen/Qwen2.5-VL-7B-Instruct, or remote with --prompt_extend_method 'dashscope' and a DASH_API_KEY. The Gradio script i2v_14B_singleGPU.py accepts --ckpt_dir_480p, --ckpt_dir_720p or both. Status at this revision: the Todo list leaves Diffusers and ComfyUI integration unchecked for image-to-video, so these commands use the project's own scripts. Architecture: flow-matching diffusion transformer with a T5 text encoder and the Wan-VAE 3D causal VAE. Licence: Apache-2.0; the card disclaims rights over generated content and forbids uses such as sharing unlawful or harmful content, misinformation or content targeting vulnerable groups. Based on the Wan-AI/Wan2.1-I2V-14B-480P card pinned to revision 6b73f84e66371cdfe870c72acd6826e1d61cf279 and observed on 2026-10-27; the 720P card at revision 8823af45fcc58a8aa999a54b04be9abc7d2aac98 matches it apart from size and paths. Nothing was run here; integration status may have changed after this revision. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P/raw/6b73f84e66371cdfe870c72acd6826e1d61cf279/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 4140:4527, 5954:6014, 6114:6704, 6706:6871; exact excerpt SHA256 d7776fec1ca16e5898e81af512c325f4c87d3a4f11a2a44d02ebd92cbfda73d0,36b99e5374e82692dfcd03d04494cf3ccf65da7ccd1a3f50cb68938b8701317f,c132f4720ec6574a07b86dd7061914957981d3496dd71d66a73dc296682e47b3,d60645bb2cde329e36"
    }
  ]
}