Commits

Commit history

1. Commits on May 26, 2022

   • make self-attention line up and configurable as in the paper

     lucidrains

     committed
     388359cView commit details

     Copy full SHA for 388359c

     View code at this pointBrowse repository at this point

   • move the resnet block towards the biggan like with norm -> activation -> projection, as in appendix

     lucidrains

     committed
     07da568View commit details

     Copy full SHA for 07da568

     View code at this pointBrowse repository at this point

   • cleanup

     lucidrains

     committed
     78acd13View commit details

     Copy full SHA for 78acd13

     View code at this pointBrowse repository at this point
2. Commits on May 25, 2022

   • offer convenient way to return list of pillow images for saving

     lucidrains

     committed
     42d87b8View commit details

     Copy full SHA for 42d87b8

     View code at this pointBrowse repository at this point

   • just make it work

     lucidrains

     committed
     c468b4fView commit details

     Copy full SHA for c468b4f

     View code at this pointBrowse repository at this point

   • remove random crop augmentation and kornia altogether, and also ready for separate noise schedules per unet in the cascade

     lucidrains

     committed
     b5deb14View commit details

     Copy full SHA for b5deb14

     View code at this pointBrowse repository at this point

   • make sure mnumber of resnet blocks are customizable, address difference between t5 encoder and google/t5-v1.1

     lucidrains

     committed
     032bbe5View commit details

     Copy full SHA for 032bbe5

     View code at this pointBrowse repository at this point

   • fix t5 to only use encoder, thanks to @jorgemcgomes again!

     lucidrains

     committed
     5a36268View commit details

     Copy full SHA for 5a36268

     View code at this pointBrowse repository at this point

   • use t5 v1.1, thanks to @jorgemcgomes for parsing the paper for this detail

     lucidrains

     committed
     b84bb7dView commit details

     Copy full SHA for b84bb7d

     View code at this pointBrowse repository at this point

   • default to t5-base

     lucidrains

     committed
     98d1b89View commit details

     Copy full SHA for 98d1b89

     View code at this pointBrowse repository at this point

   • add t5-base as usable text encoder (huggingface makes this easy)

     lucidrains

     committed
     b0fdd3dView commit details

     Copy full SHA for b0fdd3d

     View code at this pointBrowse repository at this point

   • only cosine and linear noise schedules used, just limit it to these 2, so the continuous versions are easier to build

     lucidrains

     committed
     08d3569View commit details

     Copy full SHA for 08d3569

     View code at this pointBrowse repository at this point

   • allow for config driven creation of imagen off the bat, also researcher must always give list of image resolutions this time around

     lucidrains

     committed
     2e2f87aView commit details

     Copy full SHA for 2e2f87a

     View code at this pointBrowse repository at this point

   • fix bug

     lucidrains

     committed
     fa3fde7View commit details

     Copy full SHA for fa3fde7

     View code at this pointBrowse repository at this point

   • in paper, they simply used adam, no weight decay needed, with a warning not to use adafactor

     lucidrains

     committed
     fd6763eView commit details

     Copy full SHA for fd6763e

     View code at this pointBrowse repository at this point

   • allow for overriding the lowres conditioning noise level for sampling

     lucidrains

     committed
     93e0ac9View commit details

     Copy full SHA for 93e0ac9

     View code at this pointBrowse repository at this point

   • complete the lowres augmentation level conditioning (which is basically exactly like ddpm forward pass but for the lowres conditioning image). sampling is fixed at 0.2 (in paper they proposed 0.1 o…

     Show description for 74e749a
     lucidrains

     committed
     74e749aView commit details

     Copy full SHA for 74e749a

     View code at this pointBrowse repository at this point

   • more t5 fixes

     lucidrains

     committed
     a5646e8View commit details

     Copy full SHA for a5646e8

     View code at this pointBrowse repository at this point

   • fix t5

     lucidrains

     committed
     447280eView commit details

     Copy full SHA for 447280e

     View code at this pointBrowse repository at this point

   • allow for use of larger t5, and extensible to any text encoder model

     lucidrains

     committed
     322496fView commit details

     Copy full SHA for 322496f

     View code at this pointBrowse repository at this point

   • it seems like Imagen always use film like conditioning from one of openai papers

     lucidrains

     committed
     156c496View commit details

     Copy full SHA for 156c496

     View code at this pointBrowse repository at this point
3. Commits on May 24, 2022

   • 0.0.4 release dynamic thresholding

     lucidrains

     committed
     3f6cb23View commit details

     Copy full SHA for 3f6cb23

     View code at this pointBrowse repository at this point

   • complete t5 small integration

     lucidrains

     committed
     ee79a1cView commit details

     Copy full SHA for ee79a1c

     View code at this pointBrowse repository at this point

   • basic cascading scaffold, taken from dalle2, removed anything unnecessary

     lucidrains

     committed
     f7aabebView commit details

     Copy full SHA for f7aabeb

     View code at this pointBrowse repository at this point

   • scaffold

     lucidrains

     committed
     9ca017cView commit details

     Copy full SHA for 9ca017c

     View code at this pointBrowse repository at this point
4. Loading