run_inference_from_images_cuda was routing through GPU bilinear resize on H100,
producing a different net feed and brighter output than MagicMatch (one-shot).
Co-authored-by: Cursor <cursoragent@cursor.com>
Rolling back inference_cuda alone broke GPU Full/Fast imports and hid all nodes.
CUDA build uses pipeline_cuda again; one-shot apply matches golden CPU develop stack.
Co-authored-by: Cursor <cursoragent@cursor.com>
Ports scene extract, face ONNX preprocess, and net reference resize to Torch/CUDA with minimal CPU sync, while keeping the CPU parity path unchanged.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keeps CPU parity nodes and providers unchanged; CUDA modules and ComfyUI
nodes run alongside with CPU fallback when CUDA EP is unavailable.
Co-authored-by: Cursor <cursoragent@cursor.com>
Route develop@1600 through inference createTensor resize, align reference JPEG?256?WebP net prep with the worker, and replace stale lut_hash 921fe3a5 with verified a48758ca in parity_pair.
Co-authored-by: Cursor <cursoragent@cursor.com>
Skin parse was normalizing 0-1 RGB instead of 0-255, rejecting all faces and breaking base adjustments + LUT build. Add parity_pair.py diagnostic script for local pair testing.
Probe and Photo.ts use 5000K/0 tint for JPEG sources; develop was using 5500K which skewed WB matrices and skin brightness.
Co-authored-by: Cursor <cursoragent@cursor.com>
Skip Adobe look on Standard bitmap develop, align reference WebP 256 ONNX feed and TF-style bilinear prep, and cache base adjustments from Build through Preview.
Co-authored-by: Cursor <cursoragent@cursor.com>
Include face detect/parse ONNX plus auto-WB and auto-light scene extraction so ComfyUI installs work without local conversion.
Co-authored-by: Cursor <cursoragent@cursor.com>
Move lut_encoding and render_mode to required combos with defaults, add legacy render_mode aliases, and normalize values before apply.
Co-authored-by: Cursor <cursoragent@cursor.com>
Reinitialize WebGL after canvas resize (resizing invalidates the context),
upload source at full cache resolution, and raise preview PNG cap to 1024px.
Co-authored-by: Cursor <cursoragent@cursor.com>
