- detect f2 model type for LoRAs via metadata, architecture, and filename/folder
- preprocess bare BFL-format keys with diffusion_model prefix for Flux2LoraLoaderMixin
- handle LoKR format via native NetworkModuleLokr with on-the-fly kron(w1, w2)
- add NetworkModuleLokrChunk for fused QKV split into separate Q/K/V modules
- activate native modules loaded via diffusers path
- improve error message for Flux1/Flux2 architecture mismatch
After deletion and re-clone, enso.install() runs after
install_extensions() has already passed, so install.py
was never triggered and the frontend was not built.
Generation tests cover scheduler params, color grading, and latent
corrections. Detailer tests cover model enumeration and object
detection. Both require a running SD.Next instance.
Add selectedModelId and selectedVersionId to the download bridge so
metadata JSON and preview images are fetched alongside model files.
Pass model_id and version_id from card detail view through
startCivitDownload and downloadCivitModel. Add keyword/tag mutual
exclusivity via disabled-look CSS class toggle.
Add Sort and Time period dropdowns, change Type from textbox to
dropdown. Rename Advanced accordion to Options. Layout: Model type +
Base model in one row, Sort + Time period in another. Add subfolder
toggle with template auto-fill. Wire tab-enter auto-browse with lazy
API discovery for dynamic dropdown refresh. Set Sort and Time period
values to match the initial auto-browse query. Fix duplicate elem_id
bug on tag field. Pass model_id, version_id, and base_model through
download handler to enable sidecar metadata and preview fetching.
Add base_model, model_id, version_id params to download_civit_model.
Call resolve_save_path when subfolder sorting is enabled. Pass model_id
and version_id to DownloadManager.enqueue so _fetch_sidecar downloads
metadata JSON and preview images alongside the model file. Extract
version_id from the download URL as fallback when not provided.
Relax the empty-query guard so search works with tag, sort, or period
alone. Skip exact-match name filtering when no keyword is provided so
browse results return the full API response.
replace hardcoded timestep thresholds with step-based progress percentages
so corrections work with flow-match schedulers (Flux 2, etc.)
adapt brightness, color and tint corrections for multi-channel latents:
- brightness uses multiplicative scaling instead of additive offset
- color applies to all channels instead of skipping channel 0
- tint falls back to uniform offset when TAESD encoding is unavailable
pass step parameter through correction_callback for progress calculation
- add __post_init__ to GradingParams to coerce Gradio int returns to float
- apply CLAHE to L channel in Lab space instead of per-RGB channel
- fix sharpness scaling (kornia factor 1.0=no change, map 0-based slider)
- increase shadows/midtones/highlights gamma strength 2x for visible effect
- reduce CLAHE clip slider range to 0-5 with finer 0.25 step
triton_version was initialized to False instead of None, causing
the `is None` checks to be skipped and torch_info to report triton=false
even when triton is functional
vladmandic
CalamitousFelicitousness
awsr