Stable Diffusion XL 0.9, an AI model used for generating high-quality images, has recently received an update that allows it to be used with version 1.5 models. Prior to this update, the base model of SDXL (version 0.9) had limitations in creating realistic and aesthetically pleasing faces. The faces generated using the base model appeared strange and the overall composition of the facial anatomy was off.
To overcome these limitations, a workflow was developed by taking a prompt and inputting it into the base model of SDXL along with a latent code. This generated an initial image, which was then refined by inputting it into the refiner model of SDXL. The refiner model made use of a different model than the base model, resulting in an improved image. The refinement process involved adding noise to the image to make it cleaner, followed by a face restoration step. The final result was a face that looked significantly better than the initial one.
What can version 1.5 do for your artwork
An example was provided to showcase the effectiveness of this workflow. The initial image generated by the base model of SDXL had a wonky face with strange anatomical features. However, after putting it through the version 1.5 model, the issues were resolved, and the face looked much better. This demonstrated the potential of this workflow in improving the quality of generated images.
The example workflow also mentioned an expanded workflow that used both SDXL and version 1.5 models. Although the workflow was not refined and appeared messy, it showed promise in enhancing the generated images. This expanded workflow involved taking an image from SDXL, adding noise to it, and then using a virtual version 1.5 model to further improve the image. The results were promising, with better facial and chest anatomy observed.
Comparisons between images generated using only SDXL and those generated using version 1.5 illustrated the improvements achieved. The facial anatomy, especially the eyes, looked strange in the SDXL-only images. However, once they were put through the version 1.5 model, the facial features became more defined and realistic. While there were still some issues with facial anatomy, the images generated using version 1.5 were significantly better than the base model.
In conclusion, the integration of Stable Diffusion XL 0.9 with version 1.5 models has shown great potential in improving the quality of generated images. The workflow involving the base and refiner models of SDXL, along with the addition of noise and face restoration, has led to noticeable improvements. The expanded workflow using both SDXL and version 1.5 models is still being refined but has shown promise. With further adjustments and experimentation, this integration could result in even better and more realistic generated images.
Notably, the age of the characters created with SDXL 1.0 did not correspond with the artist’s original intentions. For instance, when aiming to create a young woman in her 20s, the resulting image seemed to depict a woman in her 30s or 40s. The artist found that by using the SDXL 1.5 model, the facial quality and overall image improved significantly, rendering a more accurate representation of a youthful character.
Despite this progress, further enhancements are still required. The video example pointed out several areas where the SDXL software could be improved. One example is the representation of the iris in the eyes, which appeared distorted and unrealistic when using the SDXL 1.0 model. However, when the image was processed with the SDXL 1.5 model, the overall quality of the facial features improved, suggesting that the technology is evolving.
Another challenge the example encountered was the need to upscale the images to 4K resolution. To achieve this, they relied on a multi-diffusion upscaling technique from 1080p to 4K. While this step proved to be effective, it added an extra layer of complexity to the overall workflow.
Moreover, the example mentioned that the base model of SDXL was not usable for their purposes, as it resulted in strange-looking faces with distorted anatomical features. This limitation indicates that further refinement is needed to ensure accurate and realistic representations. Using SDXL with custom models would be more suitable once fine-tuned versions become available. Customized models will be released soon, allowing for more precise and polished artwork.
Despite these challenges and limitations, we can be hopeful about the future of SDXL. Examples of improved facial compositions and the workflow using the SDXL 1.5 model, combined with upscaling, produced desirable results. Stable Diffusion XL 0.9 shows promise as an innovative tool for AI art. Although there are still areas requiring improvement, such as facial distortion and anatomical accuracy, the advancements made in the SDXL 1.5 model are encouraging. Artists anticipate the release of fine-tuned custom models and look forward to utilizing SDXL to create stunning and realistic artwork.
excerpt: The emerging field of AI art has witnessed another significant advancement with the introduction of Stable Diffusion XL 0.9. This cutting-edge technology utilizes deep learning algorithms to generate highly realistic images of human faces. However, while there are still some shortcomings to be addressed, the results obtained from SDXL 1.5 are encouraging.
Stable Diffusion XL 0.9 has received an update that enables it to be used with version 1.5 models, resulting in improved image quality. The initial results generated using the base model of SDXL showed strange facial anatomy, but refinement through the addition of noise and face restoration using the refiner model produced significantly better results. An expanded workflow combining SDXL and version 1.5 further enhanced image quality. Future adjustments and experimentation could lead to even more realistic generated images.
1 comment