Official post-testing 3.0 release log:

Tera's got a new tiny model with unbelievable potential! Only 2 epochs, fantastic results! 💫🤍🎵

117 training images total, trained for 2 epochs (or a little over 2000 steps), with a lot of diverse depictions of her, forget about over-fitting!

3.1 release log - December 25th 2025 (Merry Christmas 🎄):

Moved for testing. There's something wrong with Tensor Art training.

3.0 release log - December 24th 2025 (happy Christmas Eve 🎄):

Moved for testing.

2.1 release log - October 20th 2025 (evening):

Improved prompt adherence and understanding of the character through custom captions, character-oriented, long and descriptive. Un-learned all the canon outfits to allow for more dynamic usage. Really, she doesn't need to learn some dress or sweater. If you want to recreate a certain look, better to just use Kontext or Qwen Edit!

2.0 release log - October 20th 2025 (morning):

Much better face details and consistency. Dynamic and flexible. Actually learned the sweater this time! 😉

1.0 for Kontext - August 9th 2025:

As a proof of concept that regular Flux LoRas are generally compatible with Kontext, I've decided to format this model differently and test how it works out.

1.0 release log - April 12th 2025:

Gemini watermark removed, model ready for use. Perfect mix of fanmade "Tera-byte" designs - from Ideogram 1.0, 2.0, 3.0 and Flux 1.0. Great in dynamic scenes and poses. Learned the iconic sweater 🤖✨

Update April 9th 2025:

Slightly lower chance of getting the Gemini watermark. Final version coming soon 🏃🏼‍♀️🏃🏼‍♀️🏃🏼‍♀️

Log April 8th 2025:

Making an "unwithered" version of Tera's Ideogram 1.0 design using Google Gemini

(Yes this model learned the Gemini watermark 🤦🏼‍♀️ I'm working on fixing that)

Android Tera-byte (OC - unwithered design i-1.0) - 3.0-2K-small