ApX logoApX logo

CogVideoX 5B

Parameters

5B

Modality

Video

Auxiliary Parameters

4.86B

Architecture

Diffusion Transformer

License

Apache 2.0

Release Date

6 Aug 2024

API Pricing (per 1M)

Self-hosted only

Hardware & System Requirements

Precision

Native

Workload Memory: 5.32 GB

INT4 · 1024x1024

DiT Backbone Weights: 2.33 GB

VAE Decode Workspace: 0.49 GB

Runtime & Graphs: 2.50 GB

Minimum Required Hardware

1x NVIDIA L4 24GB

5.3 / 21.6 GB usable

1x RTX 3060 12GB

5.3 / 10.8 GB usable

Apple M4 24GB

5.3 / 18.0 GB usable

Determine requirements for your specific workload:

Open in VRAM Calculator

Runtime requirements

Checkpoint format

FP16

CUDA

>= 11.8

PyTorch

>= 2.1

Transformers

>= 4.40.1

ArchitecturePrecisionAttention kernel

Blackwell, Hopper

Native FP16

FlashAttention-3

Ada Lovelace, Ampere

Native FP16

FlashAttention-2

Architecture Diagram

Noisy Video Latents3D Spatio-Temporal VAE Latents3D Patch Embedding3D Spatio-Temporal PatchifyHidden: 3.1k · Text: 4.86B enc · Timestep (t)x 42 DiT blocksAdaLN ModulationNorm · Timestep scale/shiftSpatio-Temporal Attention48 headsHead dim: 64+AdaLN ModulationNorm · Timestep scale/shiftFeed-Forward (MLP)Activation+Final Norm (AdaLN Scale)Predicted Velocity (v)Denoised 3D Latents to VAE

About CogVideoX 5B

CogVideoX 5B is a 3D spatio-temporal video diffusion transformer developed by Zhipu AI, utilizing a 3D VAE with 8x8x4 compression and expert transformer blocks for high-frame-rate coherence.

Specifications

Attention

Attention Heads

48

Attention Head Dimension

64

Dimensions

Base Precision

FP16

Auxiliary Parameters

4.86B

Hidden Dimension Size

3,072

Number of Layers

42

About CogVideoX


Other CogVideoX Models