Switch to full model with fp16/bf16 inference for better performance

Enable optimized SDPA attention backends for faster inference

Add NF4 4-bit inference with bitsandbytes

180s

80s

600s -> 180s

u

u

docs: update README with full project links and demo

feat: add project links with icons in header

UI improvements: move status bar to right side, simplify layout, update defaults to Wang Xizhi

fix: use float32 instead of bfloat16 for compatibility

perf: parallel loading of safetensors shards

perf: enable FlashAttention/MemEfficient SDPA backends instead of torch.compile

fix: remove flash-attn (requires source build), add batch generation support

Add flash-attn to requirements for attention acceleration

Add batch generation, torch.compile acceleration, fix dtype issues

fix: use assign=True in load_state_dict to preserve checkpoint dtype

fix: ensure consistent dtype in prepare function (use img.dtype)

fix: ensure model dtype is bfloat16 after loading safetensors

feat: support sharded safetensors for faster model loading

fix: disable 4-bit quantization for faster init on ZeroGPU

fix: increase GPU duration to 300s for model loading

fix: import spaces first and lazy load inference to avoid CUDA init

fix: add all missing dependencies from original project

fix: add opencv-python-headless

fix: pin huggingface-hub and transformers versions for compatibility

fix: version compatibility and HF_TOKEN for private repo

fix: update checkpoint path to TSXu/Unicalli_Pro

fix: shorten short_description

feat: Add UniCalli Chinese calligraphy generator

initial commit

Duplicate from gradio-templates/text-to-image-gradio-template