Source: AIGC Open Community
Image Source: Generated by Wujie AI
The Playground platform announced the open source of Playground V2, which allows commercialization. Users can generate various types of 1024x1024 images, including 3D, animation, sketch, punk, and dark through text, and also provides free online experience.
Playground V2 is developed based on Stable Diffusion XL, and collected 10 categories from Midjourney, each containing 3000 high-quality image samples for text-image alignment.
According to test data, Playground V2's generated images are much more popular than Stable Diffusion XL among over 1000 text prompts.
Free trial address: https://playground.com/
Open source address: https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic
Playground V2 is developed based on Stable Diffusion XL, so the architecture of the two is almost the same. However, the high-quality samples of 10 categories collected from Midjourney played a key role in data fine-tuning.
Playground V2 uses a larger parameter UNet as the main model, which is 3 times that of the previous Stable Diffusion model.
Additionally, multiple modules have been added, including: size and cropping coordinate conditions based on Fourier feature encoding, used to control the position of objects in the generated images.
Based on training with multiple aspect ratios, generating images with different shape proportions; for text encoding, it uses the feature splicing of CLIP ViT-L and OpenCLIP ViT-bigG. In addition, an independent detail enhancement network is used to improve the visual quality of the main model's generated images.
In addition, considering the significant differences in real image resolution and aspect ratios, researchers collected data of 20 different aspect ratios to make the model adapt to different aspect ratio training data, ensuring that the pixel count is close to 1024*1024 by switching aspect ratio data during training.
Appreciation of Images Generated by Playground V2
The images generated by Playground V2 are very good in terms of light matching, dark contrast, text description restoration, color, etc. Those who cannot use Midjourney, DALL·E 3, and Stable Diffusion can give it a try.
A snake entwined with a woman, very beautiful, watercolor painting, movie style, calligraphy lines, dark, strange, mysterious, modern retro, rich dark colors, bohemian style.
A girl and a bear, complex fur and fabric textures, digital painting, glowing effects, super fine, dramatic lighting, the girl's expression is memorable.
Swiss roll and strawberries, clean white background, realistic style, 3D effect.
Super delicious steak, movie effect, professional food photography, studio lighting, studio background, advertising photography, complex details, super fine, super realistic, 8K ultra-high definition.
The plate is filled with colorful sushi rolls, tempura vegetables, and a bowl of steaming miso soup. Fresh fish slices, visually and aromatically intoxicating. Rice vinegar, mustard sauce, pickled ginger, soy sauce, and green tea make every bite more delicious. 8K ultra-high definition, realistic.
A witch wearing a black hat, dressed in a black dress, baroque style, fashion shoot, subtle tone background, super macro, complex and realistic details, studio effect, dynamic photo, professional photo, studio photography, 8K ultra-realistic, realistic style.
A gray alien, presenting snake skin textures in different tones, integrating robot neck features, with large captivating eyes reflecting holographic effects, standing in a holographic forest swamp, wearing a Venetian iron mask decorated with Maori gold thread, 8K ultra-high definition.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
