Fix Gradio compat: remove show_download_button, replace ColorPicker with Textbox

V2: flood-fill BG removal, 20-concept parent/child detection, logo filter, color picker UI, hover download

Revert to SAM 3 mask pipeline with sigmoid logits + edge-only AA upscaler, remove rembg

Fix OMP_NUM_THREADS=3500m crash: sanitize env var before rembg import

Add onnxruntime dependency for rembg

Pipeline rewrite: SAM 3 boxes + rembg background removal + 4x Lanczos upscale

Restore sigmoid logits smooth edges + Lanczos4 upscale (the version user liked)

Smooth AA edges: narrow anti-aliased transition band instead of hard threshold

Crisp edge pipeline: separate RGB/alpha upscaling, supersampled AA threshold, stronger sharpening

Show user-friendly messages when no image uploaded, model loading, or no assets found

Add 4x Lanczos upscaler with unsharp masking for crisp high-res asset output

Add PDF batch processing: upload PDF, extract assets from every page, tabbed UI

Remove logo concept, increase min area to 500, add Download All ZIP, add asset library

Add Gradio auth: login required via APP_USERNAME/APP_PASSWORD env vars

Revert to softer edges (7x7 kernel, 2 dilation, no logit multiplier) per user preference

Sharper edges: 3x logit multiplier, 5x5 kernel, 1 dilation iteration

Force full 255 opacity inside binary mask, smooth alpha only at edge transition zone

Fix: use binary mask for area filtering, logits only for smooth alpha rendering

Fix empty tensor truthiness: check numel() instead of using or chain

Fix logits key: SAM 3 uses 'masks_logits' not 'mask_logits' - enables smooth alpha edges

Use raw logits for smooth edges: bicubic upsample + sigmoid alpha instead of binary masks

Use SAM 3 raw neural mask directly - remove polygon approximation that was degrading edge quality

Remove all blur: precise binary cutouts with morphological cleanup and contour smoothing only

Advanced 4-stage edge refinement: morphological cleanup, contour smoothing, edge-zone feathering

Smooth alpha-feathered edges, force PNG format in Gallery, remove unsupported preview param, increase crop padding

Fix Gradio 6.0 API: move theme/css to launch(), remove unsupported show_download_button

Premium UI: gradient title, larger gallery (scale=3), hover effects, download buttons, Inter font, orange theme

Output actual PNG files with transparency instead of numpy arrays (which Gradio converts to WebP)

Fix roi_align: handle tuple from .unbind() not just list

Add roi_align + layer_norm interceptors to fix HalfTensor/FloatTensor mismatch in geometry encoder

Switch to model.half() (float16) for native T4 acceleration with correct Meta checkpoint loading

DIAGNOSTIC: Remove float32 cast, let model run in native bfloat16 with interceptors. Print param dtypes

Restore full 10-concept list - safety posters need icon/image/illustration, not just chart/diagram/table

CRITICAL FIX: Pass ckpt_path to build_sam3_image_model() to prevent internal download hang without HF token

CRITICAL FIX: Remove redundant checkpoint loading that was corrupting model weights. build_sam3_image_model() already loads them correctly

Add comprehensive logging with flush=True, try/except, and reduce concepts to 3 for timeout safety

CRITICAL OPTIMIZATION: Implement aggressive gc.collect() and manual dictionary purges to prevent Linux OOM Killer RAM limits

Add debugging prints to determine why masks are empty, maybe model thresholds are too strict

CRITICAL HOTFIX: Revert to 32-bit execution to prevent float16 matrix overflows (NaN), neutralizing blank extraction masks

CRITICAL FIX: Bypass PyTorch AMP and intercept F.linear/F.conv2d to deterministically cast float matrices

CRITICAL FIX: Patch torch.utils.checkpoint to preserve float16 autocast context inside Decoder blocks

CRITICAL FIX: Wrap inference loop in explicit autocast context. Sam3Processor lacks native decorators

HOTFIX: Remove global bfloat16 alias to unbreak torch.load checkpoint parsing

CRITICAL FIX: Remove is_bf16_supported gate - T4 reports True but crashes on bf16 ops. Patches now unconditional + model cast to fp16

Intercept bfloat16 requirements globally via PyTorch module namespace aliasing

Use importlib to locate and rewrite Meta SAM 3 bfloat16 requirements across PyEnv virtual environments

Dynamically patch SAM 3 source code on T4 boot to replace hardcoded bfloat16 demands with float16 equivalents

Cast model explicitly to bfloat16 on T4 to override AMP dtype mismatch

Monkey-patch torch.autocast to natively support T4 Turing GPUs by intercepting bfloat16 demands

Protect numpy from bfloat16 errors on T4 and restrict model upcast to CPU only