v1.2.3 - easy like A-B-C

Stable Diffusion 3.5 Medium Lycoris LoKr

Bugfixes

Sana

• No more censored text embeds, eliminate refusals
• Prompt cleaning dependency fixes

Misc

• Update configure.py recommendations for multi-resolution to limit to Flux
• Update the auto-generated data backend config to be more useful and kinder to users
• Sine schedule fixed
• Missing captions = image skipped during pre-collection phase

Deprecations

• All --flux_* option names that apply more broadly to all flow-matching models have been renamed.

Pull Requests

• use correct autoencoder class for different model exports (#1245) by @bghira in #1246
• sana: remove censorship from text inputs, improve accuracy of generated embeds and eliminate refusals by @bghira in #1254
• add beautifulsoup4 for prompt cleaning in Sana by @bghira in #1260
• updates for configure.py multi-resolution suggestions & dataset cache path handling by @bghira in #1263
• rename and deprecate flux option names by @bghira in #1264
• sine schedule needs to init the optimiser states by @bghira in #1265
• skip images that have missing caption by @bghira in #1266
• update docs for new arg names + import refactored caption ignore logic by @bghira in #1269

Full Changelog: v1.2.2...v1.2.3