HOME | DD

Description

This picture was made merging in a custom model with SD1.5-ema and turned to fp16 for reduction in size. 10 weight variantians were created and one chosen. No additional material beyond what was generated with the help of the AI was used.

From this 200 iterations were created with assortments of prompts and seed with different samplers at 512x512. From the 2000 or so pictures few were selected for further refining.

Refining process has following workflow:

Gradual increase of resolution with patch of 10 choosing the best candidate. At every step undesirable elements were removed and additional desired elements introduced by painting in Photoshop; to avoid colour drifting between iterations occasionally colour, brightness, vibrance and contrast was adjusted in Photoshop. After final iterations and work in 960x960 resolution the final candidate was chosen and upscaled with BSGRAN to 2x dimension. Then finalised in photoshop by adjusting the overall colour and brightness, and cropped to removed to strange artifacting at the edges of the photo.

Repo used: Automatic1111 (commit 804d9fb) all relevant technology citations and credits can be sourced from the Github. (Which is something people should do btw. This system is built on public opensource innovation and documentation - the workflow can be achieved just about in any repo; the technology behind it all is what makes this possible and that was made by researchers and engineers and published, it was not by Anon. Which is why it is important to keep in mind that the true credit belongs to the people behind the technology and iteration of it.)

Description

This picture was made merging in a custom model with SD1.5-ema and turned to fp16 for reduction in size. 10 weight variantians were created and one chosen. No additional material beyond what was generated with the help of the AI was used.

From this 200 iterations were created with assortments of prompts and seed with different samplers at 512x512. From the 2000 or so pictures few were selected for further refining.

Refining process has following workflow:

Gradual increase of resolution with patch of 10 choosing the best candidate. At every step undesirable elements were removed and additional desired elements introduced by painting in Photoshop; to avoid colour drifting between iterations occasionally colour, brightness, vibrance and contrast was adjusted in Photoshop. After final iterations and work in 960x960 resolution the final candidate was chosen and upscaled with BSGRAN to 2x dimension. Then finalised in photoshop by adjusting the overall colour and brightness, and cropped to removed to strange artifacting at the edges of the photo.

Repo used: Automatic1111 (commit 804d9fb) all relevant technology citations and credits can be sourced from the Github. (Which is something people should do btw. This system is built on public opensource innovation and documentation - the workflow can be achieved just about in any repo; the technology behind it all is what makes this possible and that was made by researchers and engineers and published, it was not by Anon. Which is why it is important to keep in mind that the true credit belongs to the people behind the technology and iteration of it.)