CogVideoX1.5-I2V

• Testing with the diffusers library enabled all optimizations included in the library. This scheme has not been tested on non-NVIDIA A100/H100 devices. It should generally work with all NVIDIA Ampere architecture or higher devices. Disabling optimizations can triple VRAM usage but increase speed by 3-4 times. You can selectively disable certain optimizations, including:

pipe.enable_sequential_cpu_offload()
pipe.vae.enable_slicing()
pipe.vae.enable_tiling()

• In multi-GPU inference, enable_sequential_cpu_offload() optimization needs to be disabled.
• Using an INT8 model reduces inference speed, meeting the requirements of lower VRAM GPUs while retaining minimal video quality degradation, at the cost of significant speed reduction.
• PytorchAO and Optimum-quanto can be used to quantize the text encoder, Transformer, and VAE modules, reducing CogVideoX’s memory requirements, making it feasible to run the model on smaller VRAM GPUs. TorchAO quantization is fully compatible with torch.compile, significantly improving inference speed. FP8 precision is required for NVIDIA H100 and above, which requires source installation of torch, torchao, diffusers, and accelerate. Using CUDA 12.4 is recommended.
• Inference speed testing also used the above VRAM optimizations, and without optimizations, speed increases by about 10%. Only diffusers versions of models support quantization.
• Models support English input only; other languages should be translated into English during prompt crafting with a larger model.

Note

• Use SAT for inference and fine-tuning SAT version models. Check our GitHub for more details.

Getting Started Quickly 🤗

This model supports deployment using the Hugging Face diffusers library. You can follow the steps below to get started.

We recommend that you visit our GitHub to check out prompt optimization and conversion to get a better experience.

1. Install the required dependencies

# diffusers (from source)
# transformers>=4.46.2
# accelerate>=1.1.1
# imageio-ffmpeg>=0.5.1
pip install git+https://github.com/huggingface/diffusers
pip install --upgrade transformers accelerate diffusers imageio-ffmpeg

2. Run the code

import torch
from diffusers import CogVideoXImageToVideoPipeline
from diffusers.utils import export_to_video, load_image

prompt = "A little girl is riding a bicycle at high speed. Focused, detailed, realistic."
image = load_image(image="input.jpg")
pipe = CogVideoXImageToVideoPipeline.from_pretrained(
    "THUDM/CogVideoX1.5-5B-I2V",
    torch_dtype=torch.bfloat16
)

pipe.enable_sequential_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

video = pipe(
    prompt=prompt,
    image=image,
    num_videos_per_prompt=1,
    num_inference_steps=50,
    num_frames=81,
    guidance_scale=6,
    generator=torch.Generator(device="cuda").manual_seed(42),
).frames[0]

export_to_video(video, "output.mp4", fps=8)

Quantized Inference

PytorchAO and Optimum-quanto can be used to quantize the text encoder, transformer, and VAE modules to reduce CogVideoX's memory requirements. This allows the model to run on free T4 Colab or GPUs with lower VRAM! Also, note that TorchAO quantization is fully compatible with torch.compile, which can significantly accelerate inference.

# To get started, PytorchAO needs to be installed from the GitHub source and PyTorch Nightly.
# Source and nightly installation is only required until the next release.

import torch
from diffusers import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel, CogVideoXImageToVideoPipeline
from diffusers.utils import export_to_video, load_image
from transformers import T5EncoderModel
from torchao.quantization import quantize_, int8_weight_only

quantization = int8_weight_only

text_encoder = T5EncoderModel.from_pretrained("THUDM/CogVideoX1.5-5B-I2V", subfolder="text_encoder",
                                              torch_dtype=torch.bfloat16)
quantize_(text_encoder, quantization())

transformer = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX1.5-5B-I2V", subfolder="transformer",
                                                          torch_dtype=torch.bfloat16)
quantize_(transformer, quantization())

vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX1.5-5B-I2V", subfolder="vae", torch_dtype=torch.bfloat16)
quantize_(vae, quantization())

# Create pipeline and run inference
pipe = CogVideoXImageToVideoPipeline.from_pretrained(
    "THUDM/CogVideoX1.5-5B-I2V",
    text_encoder=text_encoder,
    transformer=transformer,
    vae=vae,
    torch_dtype=torch.bfloat16,
)

pipe.enable_model_cpu_offload()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

prompt = "A little girl is riding a bicycle at high speed. Focused, detailed, realistic."
image = load_image(image="input.jpg")
video = pipe(
    prompt=prompt,
    image=image,
    num_videos_per_prompt=1,
    num_inference_steps=50,
    num_frames=81,
    guidance_scale=6,
    generator=torch.Generator(device="cuda").manual_seed(42),
).frames[0]

export_to_video(video, "output.mp4", fps=8)

Additionally, these models can be serialized and stored using PytorchAO in quantized data types to save disk space. You can find examples and benchmarks at the following links:

• [torchao](https://gist.github.com/a-r-r-o-w/4d9732d17412