Skip to content

Latest commit

 

History

History
105 lines (91 loc) · 5.47 KB

NOTEBOOK-README.md

File metadata and controls

105 lines (91 loc) · 5.47 KB

🙏 Credits/Contributions

Links

Changes/Enhancements

  • July 12, 2022 13 new CLIP models and many more diffusion models have been updated in the entmike/disco-diffusion-1 GitHub fork and entmike/disco-diffusion-1 docker image. If you use RunPod, entmike/disco-diffusion-1:runpod has the models pre-baked into them as well.

    CLIP

    • ViTB32_laion2b_e16
    • ViTB32_laion400m_e31
    • ViTB32_laion400m_e32
    • ViTB32quickgelu_laion400m_e31
    • ViTB32quickgelu_laion400m_e32
    • ViTB16_laion400m_e31
    • ViTB16_laion400m_e32
    • RN50_yffcc15m
    • RN50_cc12m
    • RN50_quickgelu_yfcc15m
    • RN50_quickgelu_cc12m
    • RN101_yfcc15m
    • RN101_quickgelu_yfcc15m

    Diffusion Pixel Art (Credit https://github.com/KaliYuga-ai/Pixel-Art-Diffusion/blob/e037fd58e2aef58f28d7511ea6dcb184e898e39f/Pixel_Art_Diffusion_v1_0.ipynb)

  • July 9, 2022

  • May 28, 2022

    • Turbo Mode after frame 10 fixed. (entmike#8)

  • May 20, 2022

    • Voronoi Diagram introduced as an alternative to perlin noise and init_image as a starting point. To use, set init_generator to voronoi (Default is perlin to honor default noise behavior) voronoi_points controls how many regions the voronoi procedure creates. You can also optionally control the palette by specifying a different file in voronoi_palette. Refer to the 2 example yaml files in the new palettes folder. There are currently two modes (generated and static.) For static, you specify all your values in the yaml file as in the example in static.yaml. For randomly generated color ranges, refer to default.yaml to see how you can control what random color hue ranges will be picked.
    • Other minor bug fixes and code cleanup.

  • May 16, 2022

    • prompt_salad feature implemented. Make Mad Libs out of a prompt_salad_template!

  • May 14, 2022

    • Make this README shared across notebooks (NOTEBOOK-README.md)

  • May 13, 2022

    • Implement (very) experimental vertical symmetry. (Credit: aztec_man#3032 on Discord)
    • Add GPU detection warning for T4 and V100 GPUs. (I cannot implement the !pip install torch==1.10.2 torchvision==0.11.3 -q patch because it breaks pytorch3d, sorry guys, I'll keep trying to find another solution.

  • May 11, 2022

    • Discord link fixed

  • May 10, 2022

    • Fix, then break, then fix again, pytorch3d and 3d animations

  • May 6, 2022

    • Twilio SMS alerts (optional, disabled by default)

  • May 5, 2022

    • sqlite3 DB support to store your params and images for future query/display/searching.

  • May 4, 2022

    • Add fallback URLs to model downloads
    • Add init_images and images_out parameters to control directory locations
    • Add save_metadata (Default = False) parameter to optionally embed DD params into your .PNGs
    • Add multiplier support.

  • May 3, 2022

    • Add Symmetry Parameters
    • Modifier Support for Art Studies!

  • May 2, 2022

    • Add initial support for YAML load/export
    • Add initial logging support

  • April 2022

    • All functions moved to dd.py that are not needed in the Notebook to reduce clutter and hopefully improve readibility.

    • All other Git repos that used to get cloned and dumped in your Google Drive are now referenced as pip packages.

Command-Line Support

After running the Set Up Environment cell, from your Google Colab Terminal you can run your Disco Diffusion workload from a terminal or make a bash script to do multiple different batches. Example:

cd /content/gdrive/MyDrive/disco-diffusion-1
python disco.py --steps=50 --batch_name="CommandLineBatch" --RN50=False \
--text_prompts='{"0":["A beautiful painting of a dolphin","ocean theme"]}'

YAML Support from Terminal

Use a YAML file to save/change your settings. (See examples/configs/lighthouse.yml for an example structure.)

cd /content/gdrive/MyDrive/disco-diffusion-1
python disco.py --config_file=examples/configs/lighthouse.yml