raw/capybara.md

Raw capture — raw/capybara.md

url: https://github.com/xgen-universe/Capybara date_fetched: 2026-07-05 backfilled: true

🎉 Welcome to visit our Project Page | 💻 Visit our Demo Website to try our model!

Capybara is a unified visual creation model, i.e., a powerful visual generation and editing framework designed for high-quality visual synthesis and manipulation tasks.

The framework leverages advanced diffusion models and transformer architectures to support versatile visual generation and editing capabilities with precise control over content, motion, and camera movements.

Key Features:

🎬 Multi-Task Support: Supports Text-to-Video (T2V), Text-to-Image (T2I), Instruction-based Video-to-Video (TV2V), Instruction-based Image-to-Image (TI2I), and various editing tasks
🚀 High Performance: Built with distributed inference support for efficient multi-GPU processing
[2026.02.20]🎨 Added ComfyUI support with custom nodes for all task types (T2I, T2V, TI2I, TV2V), together with FP8 quantization support for the inference script and ComfyUI custom node.
[2026.02.17]🚀 Initial release v0.1 of the Capybara inference framework supporting generation and instruction-based editing tasks (T2I, T2V, TI2I, TV2V).
Add support for ComfyUI.
Release our unified creation model.
Release training code.

We recommend using Anaconda to create an isolated Python environment and recommend using CUDA 12.6:

# Clone the repository
git clone https://github.com/xgen-universe/Capybara.git
cd Capybara
# Create environment
conda create -n capybara python=3.11 -y
conda activate capybara
# Install pytorch (torch 2.6.0 with CUDA 12.6)
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu126
# Install dependencies
pip install -r requirements.txt
# [Optional] Install Flash Attention for faster inference
pip install flash_attn --no-build-isolation

Capybara requires the following model components:

Models	Download Link	Description
Model	Huggingface Model	All Mandatory Models
Rewrite Model	Huggingface Model	Qwen3-VL-8B-Instruct for Rewrite Instruction

Download the required models and organize them in the following structure:

ckpts/
├── scheduler/
│   └── scheduler_config.json
├── text_encoder/
│   ├── byt5-small/
│   ├── Glyph-SDXL-v2/
│   └── llm/
├── transformer/
│   └── capybara_v01/
├── vae/
└── vision_encoder/
    └── siglip/

Capybara supports two inference modes: Single Sample Mode for quick testing with a single input, and Batch Mode for processing multiple samples via CSV files. Both modes support all task types.

We provide example scripts under script/ and example data under assets/ to help you get started quickly:

assets/
├── examples/           # Example media files
│   ├── img1.jpeg
│   ├── img2.jpeg
│   ├── video1.mp4
│   └── video2.mp4
└── test_data/          # Example CSV files for batch mode
    ├── ti2i_example.csv
    └── tv2v_example.csv

Process a single image or video with a text prompt. See script/test_single_infer.sh for full examples.

Instruction-based Image-to-Image (TI2I):

python inference.py \
    --pretrained_model_name_or_path ./ckpts \
    --media_path ./assets/examples/img1.jpeg \
    --prompt "Change the time to night." \
    --output_path ./results/test_single_output/ti2i \
    --num_inference_steps 50 \
    --task_type ti2i \
    --resolution 720p \
    --rewrite_instruction

Instruction-based Video-to-Video (TV2V):

python inference.py \
    --pretrained_model_name_or_path ./ckpts \
    --media_path ./assets/examples/video1.mp4 \
    --prompt "Replace the monkey with Ultraman. Keep the Ultraman's motion matched the original running pose and motion of monkey." \
    --output_path ./results/test_single_output/tv2v \
    --num_inference_steps 50 \
    --num_frames 81 \
    --task_type tv2v \
    --resolution 480p \
    --rewrite_instruction

More inference examples about generation tasks (T2I/T2V)

Text-to-Video (T2V):

python inference.py \
    --pretrained_model_name_or_path ./ckpts \
    --prompt "A giant humpback whale and its calf gracefully swim in the crystal-clear, deep blue open ocean." \
    --output_path ./results/test_single_output/t2v \
    --guidance_scale 4 \
    --num_inference_steps 50 \
    --num_frames 81 \
    --task_type t2v \
    --resolution 480p \
    --aspect_ratio "16:9" \
    --rewrite_instruction

Text-to-Image (T2I):

python inference.py \
    --pretrained_model_name_or_path ./ckpts \
    --prompt "A group of five hikers, sitting on the snow mountain." \
    --output_path ./results/test_single_output/t2i \
    --guidance_scale 4 \
    --num_inference_steps 50 \
    --task_type t2i \
    --resolution 720p \
    --aspect_ratio "16:9" \
    --rewrite_instruction

Process multiple samples using a CSV file. See script/test_infer.sh for a full example.

For editing tasks (TI2I / TV2V), prepare a CSV with img_path/video_path and instruction columns:

img_path,instruction
img1.jpeg,instruction1.
img2.jpeg,instruction2.

The path column holds relative paths to media files (images or videos) under the data root directory.

Example CSV files are provided in assets/test_data/.

python inference.py \
    --pretrained_model_name_or_path ./ckpts \
    --csv_path ./assets/test_data/ti2i_example.csv \
    --data_root_path ./assets/examples \
    --output_path ./results/test_output/ti2i-480p \
    --num_inference_steps 50 \
    --num_frames 81 \
    --task_type ti2i \
    --resolution 720p \
    --rewrite_instruction

Use accelerate for distributed inference across multiple GPUs:

accelerate launch --config_file acc_config/accelerate_config.yaml --num_processes 2 inference.py \
    --pretrained_model_name_or_path ./ckpts \
    --csv_path ./assets/test_data/ti2i_example.csv \
    --data_root_path ./assets/examples \
    --output_path ./results/test_output/ti2i-480p \
    --num_inference_steps 50 \
    --num_frames 81 \
    --task_type ti2i \
    --resolution 720p \
    --rewrite_instruction

Capybara provides custom ComfyUI nodes for all task types (T2V, T2I, TI2I, TV2V).

ln -s /path/to/Capybara /path/to/ComfyUI/custom_nodes/Capybara| Node | Description | |---|---| | Capybara Load Pipeline | Load all model components with automatic attention backend selection | | Capybara Generate | Main generation / editing node for all task types | | Capybara Load Video | Load a video file as IMAGE frames + fps | | Capybara Load Rewrite Model | Load Qwen3-VL for prompt rewriting | | Capybara Rewrite Instruction | Expand short prompts into detailed instructions |

A sample workflow is provided in comfyui/examples/. For setup details and node documentation, see the ComfyUI README.

Task Type	Description	Input Required
`t2v`	Text-to-Video generation	`--prompt`
`t2i`	Text-to-Image generation	`--prompt`
`ti2i`	Instruction-based Image-to-Image editing	`--media_path`+`--prompt`(or CSV)
`tv2v`	Instruction-based Video-to-Video editing	`--media_path`+`--prompt`(or CSV)

Parameter	Default	Description
`--pretrained_model_name_or_path`	(required)	Path to the model checkpoint directory
`--task_type`	`tv2v`	Task type: `t2v`,`t2i`,`ti2i`,`tv2v`
`--resolution`	`None`	Output resolution: `480p`,`720p`,`1080p`
`--aspect_ratio`	`None`	Aspect ratio: `16:9`,`9:16`,`4:3`,`3:4`,`1:1`
`--num_frames`	`81`	Number of frames to generate (e.g., 81, 101, 121)
`--num_inference_steps`	`50`	Number of denoising steps
`--guidance_scale`	`1.0`	Text guidance scale for classifier-free guidance
`--num_sample_per_case`	`1`	Number of samples to generate per input
`--rewrite_instruction`	`False`	Auto-enhance prompts using Qwen3-VL-8B-Instruct
`--rewrite_model_path`	`Qwen/Qwen3-VL-8B-Instruct`	Path to the rewrite model
`--max_samples`	`None`	Limit the number of samples to process from CSV
`--quantize`	`None`	Quantize transformer weights ( `fp8`). See FP8 Quantization.

For optimal quality and performance, we recommend the following settings:

Task Type	Recommended Resolution	Recommended Steps	Note
Video(T2V, TV2V)	`480p`	`50`	Balanced quality and generation speed
Image(T2I, TI2I)	`720p`	`50`	Higher quality for static images

Notes:

Resolution: You can experiment with higher resolutions (- 1024or- 1080p).
Inference Steps: 50 steps provide a good balance between quality and speed. You can use 30-40 steps for faster generation.

Capybara supports FP8 (E4M3) weight-only quantization for the transformer via torchao. This roughly halves the transformer's weight memory, allowing larger resolutions or longer videos to fit in GPU VRAM.

Requirements:

NVIDIA GPU with compute capability >= 8.9 (Ada Lovelace or Hopper, e.g. RTX 4090, L40, H100)
torchaoinstalled (- pip install torchao)

Add --quantize fp8 to any inference.py command:

python inference.py \
    --pretrained_model_name_or_path ./ckpts \
    --media_path ./assets/examples/video1.mp4 \
    --prompt "Replace the monkey with Ultraman." \
    --output_path ./results/test_fp8 \
    --num_inference_steps 50 \
    --num_frames 81 \
    --task_type tv2v \
    --resolution 480p \
    --quantize fp8

In the Capybara Load Pipeline node, set the quantize dropdown to fp8. The node handles everything automatically -- the transformer will be loaded in FP8 on GPU while other components (VAE, text encoders, etc.) still offload to CPU as usual.

Weightsare stored in FP8 format (roughly half the memory of bf16/fp16).
Activations and computeremain in the dtype you select (bf16 or fp16). Only the weights are quantized; they are dequantized on the fly during matrix multiplications.
When FP8 is enabled with CPU offloading, the transformer stays pinned on GPU (quantized tensors cannot be moved between devices). All other models still offload normally.

This project is released under the MIT License.

This project is built upon:

HunyuanVideo-1.5 - Base Video Generation Model
Diffusers - Diffusion pipeline infrastructure
Accelerate - Distributed training/inference
SageAttention - Efficient attention mechanism

If you find Capybara useful for your research, please consider citing:

@misc{capybara2026rao,
  title={Capybara: A Unified Visual Creation Model},
  author={Rao, Zhefan and Che, Haoxuan and Hu, Ziwen and Zou, Bin and Liu, Yaofang and He, Xuanhua and Choi, Chong-Hou and He, Yuyang and Chen, Haoyu and Su, Jingran and Li, Yanheng and Chu, Meng and Lei, Chenyang and Zhao, Guanhua and Li, Zhaoqing and Zhang, Xichen and Li, Anping and Liu, Lin and Tu, Dandan and Liu, Rui},
  year={2026}
}

For questions and feedback, please open an issue on GitHub.

You can also contact us by email: zraoac@ust.hk and hche@ust.hk

⭐ If you find this project helpful, please consider giving it a star!