CogVideoX-2B
Z.ai (Zhipu AI) · China · 2024
The first open-weight video generator from the maker of CogVideoX, small enough for a 2018 gaming card and free under Apache 2.0.
CogVideoX-2B, published on 6 August 2024, was the first model of the CogVideoX line released with downloadable weights. It matters less for what it produces than for where it runs: six seconds of 720x480 video at 8 frames per second, generated on a single consumer graphics card. Using the diffusers library with all memory optimisations enabled, the producer puts the floor at 4 GB of video memory, and 3.6 GB with INT8 quantisation, which is why the project's own notes point at a GTX 1080 Ti as a working machine. The limits are hard rather than soft. The output is locked to 720x480 and the producer states plainly that no other resolution is supported, fine-tuning included. Prompts are English only and capped at 226 tokens. A clip is always 49 frames, always six seconds. Generation of one clip at 50 steps takes about 90 seconds on an A100 and about 45 on an H100 — so the cheap memory footprint is paid for in wall-clock time on smaller cards. The published weights come to roughly 1.7 billion parameters in the video transformer; the repository also carries the T5 text encoder and the 3D causal video autoencoder that the pipeline needs, so the download is heavier than the parameter count suggests. The licence is the part worth remembering. The model went out under the producer's own restrictive document and was moved to Apache 2.0 on 27 August 2024 — the same day the larger CogVideoX-5B appeared under the restrictive licence the 2B had just left. Today this is the only member of the line that anyone can use commercially without registering with the producer.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!