Becoming a prompt ai artist

Introduction

In this post, we are going to see what we can do with stable diffusion and how to get faster to satisfying results.

The tools

In order to be a prompt artist, you need the following ressources :

• A gpu installed on your computer or cloud instance (I have a GTX 1070)
• Low memory stable diffusion repo
• The prompt book which describe how to reach certain visualization
• [optional] An account on Dalle2

The gallery

Some example of generations with their original prompt.

Prompt : A_machine_to_steam_corn,retro_futurism(1979)

Prompt : A_knight_in_full_armor,_in_an_open_space,_working_on_a_computer,_with_his_manager_looking_above_his_shoulder,_low_key_lightnin

Prompt : eldritch_terrific_horror_coming_at_the_viewer,_manga,_Junji_Ito,_Kentaro_Muira,_black_and_white,_highly_detailed_on_small_deta

Prompt : a_dream_driller_tool,_close_up_on_the_drill,_4k_bokeh

Conclusion

One thing to notice, when prompt is a bit too much out of distribution, some terms start to dissapear. This can be seen on the kniwght example (I only got one example with all the key elements).

Focusing on cross overs

A lot has already been discussed and done on the subject of generation with stable diffusion

Style transfer

A_car_looking_like_Sonic_the_hedgehog,_3D_render,_trending_on_artstation

A good example of the transfer of “Sonic features” to a car.

Here I believe the car term as be understood as Car from Pixar. And the result is great

Another example of style transfer can be achieved with a predefined image.

Here a real life image is adapted to fit respect the impressionist style.

Style transfer across brands

Thomas_the_train_engine,_in_GTA_V,_cover_art_by_Stephen_Bliss,_artstation

At first, I had the impression of the GTA4 not working properly. So i tried another one.

Kermit_the_frog,_in_GTA_V,_cover_art_by_Stephen_Bliss,_artstation

Here we recognize the style of the game.

However when looking more precisely at the background, we can better understand the background of Thomas the train.

From there, it’s clear that the stable diffusion model needs to tick some boxes and desert background might tick GTA4.

Kermit_the_frog_as_a_World_of_warcraft_character,_simplistic_3D_render

Here we are looking for additional support from the last point. The next example seems to support that.

Using outpainting for better results

I want to generate a children book cover.

After several try, i find an image that I like

In the open ai interface, I have a handy way of removing what I don’t want.

The final result has its main imperfections removed.

Conclusion

I hope you will have as much fun as I did.