Update copy: remove "messy", change "chaos" to "nature"

Restore dice emoji on randomize button

Use largest mask as default (first position)

Add example images with pre-filled prompts and settings

Enable Git LFS for images

Remove text above mask gallery

Make inverted mask the default (first position)

Remove webcam/clipboard icons from image upload

Add Original Image tab for visual alignment, min slider to 25%

Align left/right columns: add frame to input, equal_height row

Change Imagination Intensity slider to 10-100% scale

Add inverted mask as 2nd gallery option, restore Raw Generation tab

Add clickable mask gallery for instant re-blending

Add blend target option: background (shape) or foreground (surroundings)

Fit entire generated image into mask shape for pareidolia effect

Fix mask blend: use full opacity instead of double-attenuating

Fix mask selection to strongly favor larger masks

Patch torch.bmm and torch.baddbmm for H200 cuBLAS, remove autocast

Disable autocast inside nn.Linear patch to prevent fp16 recast

Fix nn.Linear patch: unconditionally cast all operands to fp32

Cast fp16/bf16 nn.Linear to fp32 — cublasGemmEx also broken on H200

Patch nn.Linear to flatten >2D inputs before GEMM on H200

Patch ALL SAM batched matmul ops to fix cuBLAS on ZeroGPU H200

Avoid cuBLAS strided batched GEMM entirely for SAM attention

Fix CUBLAS bfloat16 error in SAM attention on ZeroGPU H200

Fix CUBLAS error on ZeroGPU H200

Refactor for ZeroGPU compatibility

Add opencv-python-headless for SAM

Use official Meta SAM checkpoint

Fix SAM model loading for PyTorch 2.6

Pin pydantic==2.10.6 for Gradio compatibility

Fix Gradio API schema bug

Add ZeroGPU support and fix Gradio version

Fix huggingface_hub compatibility

Fix diffusers/transformers compatibility issue

Inital commit

initial commit