CogVideoX 1.5 makes most sense when you already have a local workflow worth keeping. Its separate text-to-video and image-to-video checkpoints generate short five- or ten-second clips; the image route lets you begin with artwork you own instead of describing every detail from scratch. Diffusers inference and adaptation tools give you room to experiment, but every extra take spends GPU time, and a usable scene still needs editing. Keep it for controlled experiments or an established adaptation pipeline. For a new production setup, its compute burden and custom commercial-use terms make it a cautious choice.
Commercial use & license
The code and older CogVideoX-2B weights use Apache 2.0; the 1.5 text and image-to-video weights use the custom CogVideoX license. Commercial use requires registration and a basic commercial license. Its free commercial allowance is limited to one million service visits per month; exceeding that requires further licensing. All other license conditions still apply.
Adult content rules
The custom weight terms prohibit military and illegal uses, acts that may undermine China's national security or unity, harm society's public interest, or infringe people's rights and interests. Adult-game work must comply with those restrictions and the rights attached to source material.
What you need
Use Python 3.10–3.12 with the matching CogVideoX1.5-5B or CogVideoX1.5-5B-I2V weights and documented inference code. BF16 is recommended, and prompts should be in English. The optimized Diffusers route uses CPU offload and VAE slicing/tiling; quantization can reduce VRAM further at a speed cost. The published memory figures were tested on A100/H100 and do not guarantee performance on another GPU. Check the checkpoint's custom license before commercial use.
Ranks are recommendations for the task described, based on reviewed documentation. Model ranks cover non-explicit development work. We have not run practical benchmarks. Check content policies and commercial terms separately. How the ranks work →