Skip to main content

Behind the scences of the Startup Festivus W&B Demo

Created on November 24|Last edited on December 5
This Weights & Biases Report is part of the text to image demo for the December 2022 Startup Festivus event

Try the demo here

Using Weights & Biases for Text to Image Experimentation

Using Weights & Biases Tables we can easily log multiple different generations from a Stable Diffusion model using different generation settings and different prompts and easily explore and narrow in on the settings that work best. Below are the generations logged in W&B Tables using different settings for the different generations in the demo, e.g. the Elf generation and the Zombie Punk generations.

Elf Study 9

Prompt Used:
a comic book drawing of a beautiful elf in the snow, cute, snowing, portrait, fog, christmas, looking at camera, cute, beautiful, ice-blue eyes, icicles, kodak ektar lens
The best settings that seem to work across a variety of faces are around cfg = 15 and start_schedule = 0.3

Run: upbeat-shadow-37
1



Punk Study

Prompt:
a comic book drawing of a beautiful zombie punk rocker with a pink mohawk, portrait, tristan eaton, victo ngai, vivid, vibrant, artgerm, rhads, ross draws, colorful
  • As cfg increases and start_schedule is high, the generations become more feminine, more emphasis on the prompt (contains "beautiful" and less emphasis on the init image)
  • start_schedule of 0.45 and below doesn't allow enough creativity from the model
  • Best results are around 0.55 <= tart_schedule <= 0.6 , prefer 0.6 over 0.55 , maybe test around 0.58 and cfg on the lower side.
    • cfg 5 and 0.55 are great, use these!

Run: true-sun-15
1


APPENDIX

Santa Study

Santa2 Study

Santa3 Study

Elf Study

Elf3 Study