Fix quality loss by avoiding patch downscaling and improving text-erase crop

Add /remove_bg endpoint using rembg (u2net ONNX) for CPU background removal

Fix dependency version conflicts for Hugging Face deployment

Add /text_erase: single-call text-prompt object removal (CLIPSeg+LaMa)

Add CLIPSeg text-based segmentation endpoint /segment_text

Fix: remove duplicate torch import causing UnboundLocalError in _load_realesrgan

Align with healthy gemini-eraser-ai space: remove CLIPSeg/transformers, use RealESRGAN_x4plus.pth from GitHub

Fix: download RealESRGAN x4 from GitHub instead of COPY missing 4x-UltraSharp.pth

Fix model loading: use absolute paths, retry logic, proper model assignment

Update backend with text segmentation, UltraSharp, and lossless compositing

fix: resolve runtime error in main.py by correcting variable initialization

Optimize Real-ESRGAN CPU inference for Free Tier (1 thread, tile=128)

Upgrade backend to use Real-ESRGAN

Update backend with Vanishly rebrand and AI upscaling capabilities

feat: FSRCNN AI upscaling endpoint via OpenCV dnn_superres (PRO feature)

perf: async LaMa offload, JPEG response, server-side downscale, faster FastSAM

Optimize FastSAM inference for extremely low latency

Fix form inputs for FastAPI

Add FastSAM cloud segment endpoint

Initial commit for HF Spaces backend