Wan2.1-Fun InP and Control: variants, VRAM modes and VideoX-Fun scripts

alibaba-pai's Wan2.1-Fun family (code: github.com/aigc-apps/VideoX-Fun) has four V1.0 weights: Wan2.1-Fun-1.3B-InP (19.0 GB) and Wan2.1-Fun-14B-InP (47.0 GB) for text-and-image-to-video, trained at multiple resolutions with first-and-last-frame prediction; and Wan2.1-Fun-1.3B-Control and -14B-Control for control-video generation with Canny, Depth, Pose, MLSD and trajectory control, at 512, 768 and 1024 resolution, trained on 81 frames at 16 fps, with multilingual prompts. Setup: place weights under models/Diffusion_Transformer/<model name>/ and your own trained transformers or LoRAs under models/Personalized_Model/ for the UI. The authors verified Windows 10 and Ubuntu 20.04/CentOS with Python 3.10 or 3.11, torch 2.2.0, CUDA 11.8 or 12.1 and cuDNN 8+, on GPUs from a 12 GB RTX 3060 to an 80 GB A100, and ask for about 60 GB of free disk. The card's Docker steps still pull a cogvideox_fun image and clone the older CogVideoX-Fun repository; use VideoX-Fun instead. VRAM: every predict script has GPU_memory_mode. model_cpu_offload moves the whole model to CPU after use; model_cpu_offload_and_qfloat8 also quantizes the transformer to float8, saving more VRAM at some quality cost; sequential_cpu_offload offloads each layer after use, saving the most VRAM but running slowly. Use model_cpu_offload when VRAM allows. Scripts (per model folder under examples/): predict_t2v.py (prompt, neg_prompt, guidance_scale, seed); predict_i2v.py with validation_image_start as the first frame and validation_image_end as the last frame; predict_v2v.py with validation_video; predict_v2v_control.py with a control_video already processed by Canny, Pose or Depth; lora_path in the predict scripts adds a LoRA. app.py starts a Gradio UI with the same modes, and ComfyUI is documented in comfyui/README.md. Licence: Apache-2.0. Based on the alibaba-pai/Wan2.1-Fun-1.3B-InP card pinned to revision 9bcbbcf673c0e6a996df6f3f25bee596ffa36d47 and observed on 2026-10-27; the Chinese README was read directly. Nothing was run here; the usage walkthrough uses CogVideoX-Fun as its example, and the exact Wan2.1-Fun example folder and default settings were not inspected. 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/alibaba-pai/Wan2.1-Fun-1.3B-InP/raw/9bcbbcf673c0e6a996df6f3f25bee596ffa36d47/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 807:2304, 1323:2304, 7089:7145, 8151:8211, 9316:9427, 10468:12067; exact excerpt SHA256 69c1579955c81b5a099af044b86a934786d011678bf9541a98b2be48594345cf,19e011b74a5b6e02313ed4c67f36390ea98b81031c82610a7bdb37aa6c9d0320,718d08ea3ce6ef8a082abaf25b3300e83e3a3c78627463d47714533e4903"
    }
  ]
}