Deep Analog converts digital photos to film in real time
A new AI model turns any digital photo into vintage film — instantly, with no retraining.
Yitong Mu from the Rochester Institute of Technology has just dropped a breakthrough in AI-powered film emulation. In a new arXiv paper titled *Deep Analog: Open-Set Film Emulation with Reference-Conditioned 3D LUTs* (arXiv:2608.14702), Mu tackles the long-standing challenge of converting digital images into the aesthetic of analog film — but crucially, without being limited to a predefined set of film stocks.
Mu’s system, Deep Analog, introduces StyleLUTNet, a neural model that predicts a single 3D LUT (lookup table) directly from a reference film frame. Unlike prior approaches that rely on fixed LUT banks or reconstruction-only training, StyleLUTNet uses self-supervised learning on procedurally generated color transforms. The result: it generalizes to *unseen* film stocks without requiring paired data or retraining. Deep Analog combines this color backbone with histogram-based tone matching and a physics-informed optical renderer that simulates multi-scale film grain and per-channel halation. On 350 self-supervised image pairs, the model hits 22.05 dB PSNR and 0.925 SSIM in the color stage, and 21.72 dB / 0.923 with the full pipeline — all while processing 1080p images in just 5.2 milliseconds (192 FPS). It even exports portable .cube LUTs compatible with standard editing tools like Photoshop and DaVinci Resolve.
- StyleLUTNet by Yitong Mu (RIT) predicts a 3D LUT directly from one reference film frame for open-set film emulation
- Achieves 22.05 dB PSNR / 0.925 SSIM at 192 FPS (5.2 ms/1080p) and exports standard .cube LUTs
- Generalizes to unseen film stocks without paired data or retraining using self-supervised color transform learning
Why It Matters
Replaces manual LUT hunting with one-click, reference-driven film grading — finally making analog aesthetics scalable for digital creators.