fix: restore flash-attn 2.8.3 — hard import in trellis attention module

fix: remove flash-attn — caused RunPod xformers conflict, not required by either env

docs: add inline documentation to all modified files

fix: use cu128 wheels matching ZeroGPU's torch CUDA 12.8 environment

fix: build diff_gaussian_rasterization from source for H200 (sm_90) GPUs

fix: disable NVML checks for ZeroGPU virtualized GPU environment

fix: add antialiasing param and handle 3-value return from rasterizer

fix: adapt gaussian renderer to standard diff_gaussian_rasterization API

fix: replace wheels with torch 2.8.0 ABI-compatible builds

fix: xformers 0.0.32 → 0.0.32.post2 (base version doesn't exist on PyPI)

fix: upgrade torch 2.4→2.8.0 for ZeroGPU compatibility

Update app.py

Update app.py

Pin fastapi and starlette to fix Gradio 4 compatibility

Update GPU duration from 120s to 60s

fix: handle non-RGB images in BiRefNet (CMYK, grayscale, LA modes)

fix: remove gr.api() calls incompatible with installed Gradio version

fix: add preprocessed_images to preprocess_multi; auto-detect multiimage mode from data in image_to_3d

Bump max_hole_size to 0.06 - middle ground between 0.04 (too weak) and 0.1 (too aggressive)

Dial back hole-filling: revert max_hole_size to 0.04, modest nbe increase with floor of 96, keep simplify connection

Improve hole-filling: higher thresholds, min nbe floor of 128, double-pass fill

Add warning comments for ZeroGPU limits

Update GPU durations to 120s for ZeroGPU free tier

Revert GPU durations to 180s to conform to ZeroGPU bounds

Update generation and export duration to 240s

Fix CUDA runtime error on Stateless GPU ZeroGPU environment

fix: add kornia+timm deps, use utils_birefnet.py wrapper (same as ASM-o2.1-MULTIIMAGE)

feat: use BiRefNet for background removal in both single and multi-image paths

fix: patch _json_schema_to_python_type to handle True/False bool schemas

fix: full Gradio 5.x compatibility - replace deprecated gr.update() with component returns

fix: monkey-patch gradio_client bool schema bug directly at runtime

fix: pin gradio==4.44.1 for ssr_mode support, remove runpod dep

Fix iPhone image orientation bug by adding pillow_heif

fix: restore ssr_mode=False to prevent Gradio api_info() crash on startup

fix: resolve build conflict + 150s export timeout

Update gamma correction to 1.8

Fix float type error in sampler steps

Revert "Fix: Bypass C0 multiplication for DC-only SH - directly add 0.5 bias for correct shadow rendering"

Fix: Bypass C0 multiplication for DC-only SH - directly add 0.5 bias for correct shadow rendering

Fix: Change max_sh_degree to sh_degree for Python SH conversion

Fix: Apply proper DC bias fix - Enable Python SH conversion with +0.5 bias, remove masked boost patch

Revert: Remove all boosts and gamma correction - Restoring original behavior

Fix: Remove division by C0 (0.282) - SH boost amount is now RAW +0.5

Fix: Masked Constant Boost (+0.5) - Exact shadow color match, no variable curve

Fix: Restoration of Smart SH Boost (Threshold 0.4, Max 0.7) - Overwriting regression

Fix: Update preprocess_image to return 2 values for UI debug

Fix: Revert aggressive Hump boost, use Soft Boost for natural colors. Add Debug Image to UI.

Fix: Safe SH Boost - Hump curve to preserve True Black (0.0) while lifting shadows

Fix: Smart SH Boost - Only lift shadows (<0.4 lum), preserve highlights

Fix: Remove global gamma, bake BiRefNet model, optimize Docker for startup