* feat: add performance loras to the end of the loras array
* fix: resolve circular dependency for unit tests
* feat: allow multiple matches for each token, optimize and extract method cleanup_prompt
* fix: update unit tests
* feat: ignore custom wildcards
* add nsfw image censoring
activatable via config, uses CompVis/stable-diffusion-safety-checker
* fix progressbar call for nsfw output
* use config to set cache dir for safety checker
* add checkbox black_out_nsfw
makes both enabling via config and checkbox possible, where config overrides the checkbox value
* fix: add missing diffusers package
* feat: extract safety checker, remove dependency to diffusers
* feat: make code compatible again after merge with main
* feat: move censor to extras, optimize safety checker file handling
* refactor: rename folder safety_checker_models to safety_checker
adapted code from https://github.com/comfyanonymous/ComfyUI/pull/3370
TODO: check if virtual scheduler tcd is needed for using sampling_base ModelSamplingDiscreteDistilled or if it's better to use sgm_uniform directly without patching
* Update async_worker.py
* Update private_logger.py
* refactor: only show full prompt details in logs, exclude from image metadata
---------
Co-authored-by: Manuel Schmid <9307310+mashb1t@users.noreply.github.com>
Co-authored-by: Manuel Schmid <dev@mash1t.de>
* Revert "fix: use LF as line breaks for Docker entrypoint.sh (#2843)" (#2865)
False alarm, worked as intended before. Sorry for the fuzz.
This reverts commit d16a54edd6.
* feat: add VAE select
* feat: use different default label, add translation
* fix: do not reload model when VAE stays the same
* refactor: code cleanup
* feat: add metadata handling
* fix: load image number from preset (#2611)
* fix: add default_image_number to preset handling
* fix: use minimum image number of preset and config to prevent UI overflow
* fix: use correct base dimensions for outpaint mask padding (#2612)
* fix: add Civitai compatibility for LoRAs in a1111 metadata scheme by switching schema (#2615)
* feat: update sha256 generation functions
29be1da7cf/modules/hashes.py
* feat: add compatibility for LoRAs in a1111 metadata scheme
* feat: add backwards compatibility
* refactor: extract remove_special_loras
* fix: correctly apply LoRA weight for legacy schema
* docs: bump version number to 2.3.1, add changelog (#2616)
* feat:support download huggingface files from a mirror site
---------
Co-authored-by: Manuel Schmid <9307310+mashb1t@users.noreply.github.com>
* fix problem
1. In partial redrawing, when refiner is empty, enable use_synthetic_refiner. The default switching timing of 0.5 is too early, which is now modified to SDXL default of 0.8.
2. When using custom steps, the calculation of switching timing is wrong. Now it is modified to calculate "steps x timing" after custom steps are used.
* fix: parse width and height as int when applying metadata (#2452)
fixes an issue with A1111 metadata scheme where width and height are strings after splitting resolution
* fix: do not attempt to remove non-existing image grid file (#2456)
image grid is actually not an image here but a numpy array, as the grid isn't saved by default
* feat: add troubleshooting guide to bug report template again (#2489)
---------
Co-authored-by: Manuel Schmid <9307310+mashb1t@users.noreply.github.com>
Co-authored-by: Manuel Schmid <manuel.schmid@odt.net>
Manuel Schmid
Manuel Schmid
cantor-set
Manuel Schmid
docppp
delta_lt_0
xhoxye