self-hosted/ai
§01·model · /models

CogVideoX 1.5

videoactiveCogVideoX License

5B text-to-video DiT by THUDM/Zhipu, generating up to 1360×768 through diffusers — CogVideoXTransformer3DModel with T5 and a 3D causal VAE, BF16 from about 10 GB and INT8 via torchao from about 7 GB. Six cards carry a recipe here: the two 12 GB 4070s set the floor and lean on offload, the 24 GB cards run the full 1360×768, and the 5090 is the only one that drops sequential offload altogether. CogVideoX License.

§02·GPUs that run this model
6 total

benchmarked·~ runs via recipe (not benchmarked)· untested·doesn't fit