ToolTulip image upscaler — provenance and redistribution notices Checked 2026-10-17. No third-party rights or legal warranty is implied. LEARNED WEIGHTS (separate from runtime and wrapper licensing) ONNX Model Zoo, Super Resolution / Sub-Pixel CNN 2016. The weights distributor's MODEL-CARD.md explicitly declares: license: apache-2.0 model_name: super-resolution-10.onnx Pinned model card and unmodified learned graph/weights: https://huggingface.co/onnxmodelzoo/super-resolution-10/tree/395472c4b6b92f788357237983d83fd913cfe7e3 File super-resolution-10.onnx, 240,078 bytes, SHA-256: 85f36ff88cc504a24af5e0602148bc56a8aa09a58eca8c0da2756f3e8186035e Apache 2.0 permits commercial redistribution subject to its conditions. This is an explicit model-card grant, not an inference from a wrapper's license. Apache license text: MODEL-LICENSE.txt, from ONNX Model Zoo source revision: https://github.com/onnx/models/tree/c5cae0f1942c8d7727372c7b1f6b485aa39e4421 Model source, model card and reference inference notebook: validated/vision/super_resolution/sub_pixel_cnn_2016/ Training reference: https://github.com/pytorch/examples/tree/main/super_resolution The model card states training on crops from the 200 BSD300 training images. No training images or benchmark fixtures are redistributed with the web tool. The learned parameters are unchanged. No training/fine-tuning was performed. The checked model directory has no additional upstream NOTICE. ACTUAL GRAPH AND PRE/POSTPROCESSING Original author paper: Shi et al., Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network. Float32 input input [1,1,224,224], output output [1,1,672,672], opset 10. The card's dynamic-size description is incorrect for this particular graph: real CPU inference rejects other input sizes. Fixed 224-square tiles are used. Native scale is THREE, not two. Final TWO-times output resamples native learned Y by half-pixel bilinear sampling. This is disclosed in the UI. Full-range (not studio/video-range) YCbCr: Y=.299R+.587G+.114B, float32 Y/255. Unlike the card's PIL example we do not quantize Y to a byte before inference, or distort the whole source to 224 square. Cb/Cr are conventionally bilinear interpolated, combined with predicted Y, then RGB is clamped to [0,255]. No normalization mean/std or video-range 16/235 scaling is applied. Predictions are finite-checked and clamped. Edge-replicated halo 16, core 192, overlap 16; tapered blending in output luminance. The model convolution receptive field is 11 source pixels: 5x5 then three 3x3 convolutions. Halo exceeds its radius and resampling support. Alpha uses premultiplied bilinear interpolation, not neural prediction. Within an 8-source-pixel neighbourhood of any nonopaque sample, RGB also uses conventional premultiplied bilinear interpolation to prevent matte halos. The inference-only RGB matte is white; it is not exported. No guarantee of recovered true detail, identity, legible text or seam-free output. RUNTIME (reused files, no duplicate WASM distribution) ONNX Runtime Web 1.30.0, copyright Microsoft Corporation, MIT. Source pin f2c39fe2f838cf35ce7da92824f5a5e3ee6e88a7: https://github.com/microsoft/onnxruntime/tree/f2c39fe2f838cf35ce7da92824f5a5e3ee6e88a7 Full MIT license and native third-party notices already distributed at: ../background-remover/ORT-LICENSE ../background-remover/ORT-ThirdPartyNotices.txt Runtime assets reused unchanged from /vendor/background-remover/: ort.wasm.min.mjs (50,126 bytes) 219e6a1fc8a9938268d18efca3c91d310bd2f4a59bbd13744df5b2b7fc6cee3b ort-wasm-simd-threaded.mjs (24,381 bytes) e13f7f94fc51b4ca72b12faeb1ee95f4ace6dfbc8939bc718aabdc0a27c4299b ort-wasm-simd-threaded.wasm (14,239,897 bytes) 3398c10d07d229bd91b364548e130e0e51a8e5704b88c7c083ebbeb78842dee2 Single-thread CPU SIMD WASM only, no WebGPU/COOP/COEP requirement. The model cache contains its four fixed verified assets, never user data. Clear cache removes this tool's cache, not background-remover's separate cache. Download total 14,554,482 bytes; largest file below 15 MB. No chunking required. No hosting/CSP/header changes. Existing .mjs and .wasm MIME rules are reused. Final total static-site size and hosting acceptance remain batch-release gates.