︎
Work                  Events                 About

RIVEN RATANAVANH


Making Visual Art with GANs

Week 3

CLIP (Text-to Image) and Style Transfer

Notes:
CLIP
- Prompt engineering is important
- SEO becomes a part of the model
- Trained in 2019 (so it knows nothing about what happened after that)
One of the challenges will just be deciding what to do since there are so many possibilities.

Prolly how 3D models will be modified a year or so from now.


CLIP Experiments
Using Hypertron V.2

Prompt: “A realistic human flower, by Nicolï Genock. Clouds and fog.:1 | Trending on ArtStation, made of mist. 4K HD, realism.:0.63”

510
530
710
960


690 was probably my favorite





A person dancing, by Nicolï Genock. Clouds and fog. Many hands.:0.6 | Trending on ArtStation, made of mist. 4K HD, realism.:0.63


120


210


350

440



A ballerina, by Nicolï Genock. Clouds and fog. Many hands.:1 | Trending on ArtStation, made of mist. 4K HD, realism.:0.63

120


600



A beautiful portrait, by Vandera La Ciana. Clouds and fog. Hands.:1 | Trending on ArtStation, made of mist. 4K HD, realism.:0.63


240


630

970




A bust, by Vandera La Ciana. Clouds and fog. A slender arm.:3 | Trending on ArtStation, made of mist. 4K HD, realism.:0.63

1160


A silhouette, by Vandera La Ciana. Clouds and fog. Flowers.:3 | Trending on ArtStation, Vray tracing. 4K HD, realism.:0.63

810

940



A bouquet of flowers, by Vandera La Ciana. Clouds.:3 | Trending on ArtStation, made of mist. 4K HD, realism.:0.63

130


320



A painting of flower
***


Style Transfer Experiments
Using this.

Actually think of Style Transfer more as texture transfer than style transfer.

This produce a fairly clear, straighforward result:

Content

Style

Result





This combo and result is kind of cursed:

Content

Style

Result:




Cleaned up the style file by cropping (could have chosen a better image of veins):

Content

Style



This one just looks like a pencil drawing, I think:



Bit hit and miss, but can adjust layers further.







(Note on modifying code:)



!python neural_style.py --content_img /content/content2.jpeg --style_imgs /content/style.jpeg

to:

!python neural_style.py
--content_img /content/content2.jpeg
--style_imgs /content/style.jpeg
--max_iterations 500 --style_scale 0.5