Wan2.1-I2V-14B-720P
Alibaba Cloud · China · 2025
The model that put open image-to-video at 720p: hand it one photograph and it animates that exact frame, at high definition, from weights anyone can download.
Wan2.1-I2V-14B-720P is the high-definition image-to-video member of the Wan 2.1 family, published together with the rest of the generation on 25 February 2025 under Apache 2.0. Where the text-to-video models invent a scene from a written description, this one is handed a photograph and asked to keep it: the first frame of the output is the picture the user supplied, and the model's job is to continue it plausibly in motion. That constraint changes the architecture. The configuration file shows a 40-layer diffusion transformer of width 5120 with 40 attention heads, the same backbone the text-to-video 14B uses, but the input width is 36 channels instead of 16 - the extra channels carry the encoded conditioning image and its mask alongside the video latent. A separate image encoder, an open-source CLIP built on xlm-roberta-large and a ViT-huge vision tower, reads the photograph; the shared umt5-xxl encoder reads the text prompt, with a ceiling of 512 tokens. A note for anyone reading the name literally: the weight files in this repository count 16.4 billion parameters, not 14. The label describes the generator shared with the text-to-video model; the surplus is the machinery that makes conditioning on an image possible. The download is correspondingly heavy - the transformer shards plus an 11.4 GB text encoder, the CLIP vision encoder and the Wan video autoencoder. One practical detail is easy to miss and matters in use. For image-to-video the size setting describes the area of the generated clip, not its shape: the aspect ratio follows the input photograph. A portrait picture yields a portrait clip without any further instruction. Alibaba split the image-to-video job across two separate downloads, this one trained for 1280x720 and a sibling trained for 832x480, and the split shows in how people use them. The 480p model is pulled from Hugging Face roughly two and a half times more often than this one, while this one has been bookmarked by nearly three times as many people - admired at the higher resolution, run at the lower one. The repository also lacks the LICENSE.txt file its card refers to, unlike the 480p sibling, where the file is present.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!