The custom-transformer apply (DiT branch) and the revert-to-default path now call reload_model_weights(force=True) instead of load_diffuser, so both go through the same managed reload: unload-before-rebuild for a lower VRAM peak and consistent job/checkpoint bookkeeping.
Gate the cross-arch unet reset on an actual checkpoint change so a same-checkpoint reload skips the pipeline-class comparison, which could false-positive and clear the override when the model was left in an img2img/inpaint variant after an interrupted generation.
load_unet returned early when sd_unet was set to 'Default'/'None',
leaving any previously-loaded custom UNET in memory. The model stayed
on the override and the dropdown looked stuck.
Now detects the transition from a custom selection to Default and
triggers sd_models.load_diffuser() to reload via the canonical
from_pretrained path.
Anima community files bundle the llm_adapter inline alongside the
transformer, so CosmosTransformer3DModel.from_single_file mangles the
118 adapter keys via Cosmos 2.0 substring renames. New helper reads the
safetensors, strips the BFL-style prefix variants, partitions adapter
keys out, runs the rest through diffusers' converter, and loads each
half into its target. SDNQ and layerwise quantization honored; TRT and
GGUF skipped.
Nunchaku's SDXL UNet does not support offloading and raises
NotImplementedError when offload=True is passed. Skip the parameter
for SDXL and log a warning instead of crashing.
- Rename HuggingFace org from nunchaku-tech to nunchaku-ai across all
nunchaku model repos (flux, sdxl, sana, z-image, qwen, t5)
- Add per-torch-version nunchaku version mapping instead of single global
version, with robust torch version parsing
Vladimir Mandic
CalamitousFelicitousness
awsr
Disty0