This folder contains PanGu Draw 3.0 models implemented with MindSpore.
In contrast to version 2.0, Pangu Draw 3.0 has been subject to experimentation and updates across various aspects, including multi-language support, diverse resolutions, improved image quality, and model scaling. This includes:
- The current industry's largest 5-billion-parameter Chinese text-to-image model.
- Supports bilingual input in both Chinese and English.
- Supports output of native 1K resolution images.
- Outputs images in multiple size ratios.
- Quantifiable stylized adjustments: cartoon, aesthetic, photography controller.
- Based on Ascend+MindSpore for large-scale training and inference, using a self-developed MindSpore platform and Ascend 910 hardware.
- Utilizes self-developed RLAIF to enhance image quality and artistic expression.
Dec 12, 2023
Support inference of PanGu Draw 3.0 model for text-to-image generation.
Please make sure the following frameworks are installed.
- python >= 3.7
- mindspore >= 2.2.10 [install]
Install the dependent packages by running:
pip install -r requirements.txt
The text-to-image task of the Pangu model requires pre-training parameters for both the low timestamp model and the high timestamp model (The pre-training model parameters is coming soon).
Version | MindSpore Checkpoint |
---|---|
Pangu3-low-timestamp-model | pangu_low_timestamp-127da122.ckpt |
Pangu3-high-timestamp-model | pangu_high_timestamp-c6344411.ckpt |
After obtaining the weights, start the demo using:
# run txt2img on Ascend
export MS_PYNATIVE_GE=1
python demo/pangu/pangu_sampling.py \
--device_target "Ascend" \
--ms_amp_level "O2" \
--config "configs/inference/pangu_sd_xl_base.yaml" \
--high_solution \
--weight "path/to/low_timestamp_model.ckpt" \
--high_timestamp_weight "path/to/high_timestamp_model.ckpt" \
--prompt "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
Note: sampled 40 steps by PanGu Draw 3.0 on Ascend 910*.
Fig1: "一幅中国水墨画:一叶轻舟漂泊在波光粼粼的湖面上,舟上的人正在饮酒放歌"
Fig2: "坐在海边看海浪的少年,黄昏"
Fig3: "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"