CogVideoX 5B on a Dedicated GPU

CogVideoX 5B from THUDM is a 5B-parameter text-to-video model that produces 6-second 720p clips with better quality than smaller alternatives. On our dedicated GPU hosting it needs a 24 GB+ card for comfortable inference.

Contents

• VRAM
• Deployment
• Performance
• Image-to-video variant

VRAM

~22 GB at FP16 for 5B base. ~12 GB at FP8. CPU offload can run it on smaller cards at significant speed cost.

Deployment

from diffusers import CogVideoXPipeline
import torch

pipe = CogVideoXPipeline.from_pretrained(
    "THUDM/CogVideoX-5b",
    torch_dtype=torch.bfloat16,
).to("cuda")

video = pipe(
    prompt="A timelapse of clouds forming over a mountain valley",
    num_inference_steps=50,
    num_frames=49,
    guidance_scale=6,
).frames[0]

6 seconds of video at 720p requires the full 50-step schedule for best quality. Reducing to 30 steps saves time with a small quality cost.

Performance

Image-to-Video

CogVideoX ships an image-to-video variant (CogVideoX-5b-I2V) that takes a starting frame plus a motion prompt. Output coherence is noticeably better than pure text-to-video. Recommended for product workflows.

See LTX Video (faster alternative) and Hunyuan Video.