v2.3.1

What's Changed

• Improved handling of stripping off eos/bos in cases where padding tokens have additionally been inserted after the eos by @damian0815 in #119
• Tokenization info, CLS token handling options for longer prompts, bugfixes by @damian0815 in #120
• Fix "sequence length 78" issue with sdxl by @damian0815 in #126
• Add device option for convenience wrapper. by @dx2-66 in #123
• v2.3.1 by @damian0815 in #127

New Contributors

• @dx2-66 made their first contribution in #123

Full Changelog: v2.3.0...v2.3.1

v2.3.0

PR #120 Tokenization info, CLS token handling options for longer prompts, bugfixes

Full Changelog: v2.2.1...v2.3.0

v2.2.1

What's Changed

Fix for #116

v2.2.0

What's Changed

• Add support for Flux by adding internal support for T5TokenizerFast & T5EncoderModel. This probably also enables support for other T5-based models - I have only tested Flux.
• Add CompelForSD, CompelForSDXL, and CompelForFlux classes to help wrangling the different embeddings and prompt intentions (inputs - main, style, negative main, negative style; outputs - embeds, negative embeds, pooled embeds, pooled negative embeds)
• Performance improvement when no weights are applied by bypassing Compel's weighting. Causes slightly different output with FP16 - disable with compel.disable_no_weights_bypass()

2.1.1

Fix RuntimeError: token sequence mismatch + default to SENTENCE splitting

Full Changelog: 2.1.0...2.1.1

v2.1.0

What's Changed

• Update README.md by @GerbertBless in #101
• Add better options for splitting long prompts by @damian0815 in #112

New Contributors

• @GerbertBless made their first contribution in #101

Full Changelog: v2.0.3...2.1.0

v2.0.2

i'm going to try to do git releases. here's the "first".

compel 2.0.2 is now available on pypi:

pip install compel==2.0.2

it fixes an issue with SDXL if you have called enable_sequential_cpu_offload() on your pipeline. you'll need to pass eg device='cuda' to compel's __init__.

i also cleaned up the SDXL demo notebook and extended it to demonstrate using positive+negative and long prompt support with the __call__ interface (i.e. compel([positive_prompt, negative_prompt])).