final output

Fix: Split Image and 3D pipeline initialization to prevent OOM on ZeroGPU

Fix: text2image_fn return count and pipeline init logic

Optimize: Default to 'a red car' and n_sample=1 for faster demonstration

Fix: Correct import path for download_kolors_weights

Fix: move weight downloads to startup to prevent ZeroGPU timeouts during generation

Fix: complete lazy loading refactor for all models in common.py

Fix: use map_location='cpu' in AestheticPredictor to avoid CUDA errors at startup on ZeroGPU

Fix: implement lazy loading for CUDA pipelines to support Hugging Face ZeroGPU

Fix: UnboundLocalError in set_attention_backend when GPU is not available during import

Fix: quote python_version to avoid interpretation as 3.1

Fix: specify python_version 3.10 to match wheels

Resolve README.md conflict and switch to Gradio SDK

initial commit

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

Initial clean commit