Skip to main content

· 3 min read

We are hosting a competition where the community can showcase their most inventive use of textual inversion concepts in text-to-image or text-to-video.

Our compute cluster; Nataili, currently comprises of 3 nodes, two have 3090, the other has 2 x A5000.

We estimate Nataili can handle 12 concepts per hour, and we can add more workers if there is high demand.

Hopefully demand will be high, we want to train hundreds of new concepts!

Schedule

2022/10/20 - Stage 1 begins, train concept command opened for usage

2022/10/22 12AM UTC - Stage 2 begins, text to image command opened for usage

2022/10/22 12PM UTC - Stage 1 ends, train concept command closed

2022/10/24 12PM UTC - Stage 2 ends, no more entries will be accepted

2022/10/24 6-12PM UTC - Winners announced

What does most inventive use mean?

Whatever you want it to mean! be creative! experiment!

There are several categories we will look at:

  • anything that's particularly creative, ~ artistic ~ or a e s t h e t i c

20221019203426_00000

  • composition; meaning anything related to how big things are, their position, the angle, etc

  • styling;

image

  • The Sims(TM): Stable Diffusion edition

So I can trai-​

  • Yes, as long as it's sfw

The Sims(TM): Stable Diffusion edition ?​

For this event the theme is “The Sims: Stable Diffusion edition”.

So we have selected a subset of products from Amazon Berkely Objects dataset.

Any other object is welcome too these are just a good source of data for this part of the competition.

Each product has images from multiple angles, the train concept command accepts up to 10 images, so choose the angles and modify backgrounds, experiment!

The goal with this category is to generate an image using the trained object, and the other categories apply, your imagination is the only limit! style a couch, try to make a BIG couch, try to make a couch on top of a mountain, try to make a vaporwave couch, anything!

How do I train a concept using the discord bot?

Type /trainconcept then press tab to go through the fields

Concept name is just a name for your concept, it doesn't have to be a single word

Placeholder is what you will use in prompts to represent your concept Add < and > so it is unique, multiple words should be hyphenated

Initializer is used as the starting point for training your concept, so this should be a single word that represents your concept

Minimum 2 images. Squareish aspect ratios work best

Untitled-2 image

How do I use the trained concept?

Prompting with concepts​

When your concept is trained you can use it in prompts.

a cute <nvidiafu> as an astronaut:

image

or a green <green-couch> sitting on top of a floor, a 3D render, trending on polycount, minimalism, rendered in cinema4d:

image

Using concepts in the webui​

The discord bot will give you a link to a .zip file, download this, extract it, and put the folder in stable-diffusion-webui/models/custom/sd-concepts-library

image