Last night, the Dream model quietly conducted a gray test on the image 3.1 model. Compared to the 3.0 version, the images generated by the 3.1 model have stronger cinematic and narrative feelings, with richer scenes. It also responds better to some artistic prompts.
For example, in the case of a close-up photo of a little girl in a supermarket, the 3.0 model may have a stronger "mannequin" feel. Whether it's the details on the skin or the overall environmental atmosphere, lighting layers, etc., the performance of 3.1 is much better.
In terms of artistic style, the 3.1 model also performs better. Compared to the 3.0 model, the 3.1 model has more accurate control over the style and clearer expression of visual features.
For example, AIBASE input the following prompt: Art Deco, oil painting, beautiful woman, long flowing hair, wavy hair, delicate face, wearing an elegant and light pink long-sleeved dress. She holds a bouquet of colorful flowers, including yellow, blue, and white flowers. The background is a soft abstract mix of light blue and white.
From the comparison of the generated images, it can be seen that the 3.1 model has a more accurate understanding of the style, and the details and lighting effects are more in line with the requirements.
On the detail level, the images generated by the 3.1 model are more realistic, such as the pores on the skin, the texture of the hair, and the realistic material texture of the clothes.
The comparison of images generated by the same prompt is as follows:
Prompt: Close-up, a girl with a white veil
However, for users who require high consistency in generation, the 3.0 model might still be a better choice. After the update, the differences between images generated in the same batch by the 3.1 model are quite significant, and the probability of repeated composition for images generated from one prompt is generally lower.
The 3.1 model is still in gray testing, and it is expected to be fully released soon.
