Deep Analog: Open-Set Film Emulation with Reference-Conditioned 3D LUTs

Deep Analog tackles a familiar problem in color tooling: fixed LUT systems are fast, but they only cover the looks they were trained on. The work frames real-time enhancement as a gated mixture-of-experts problem and shows that, without extra balancing, the gate tends to collapse onto one LUT, wasting the rest of the bank.

To get around that closed-set limit, the pipeline switches to a reference-conditioned design called StyleLUTNet. Instead of blending from a fixed LUT library, it predicts a single 3D LUT as a residual from one example frame, then trains it with self-supervision on procedurally generated color transforms. That lets the system generalize to unseen film stocks without paired data or retraining.

The full Deep Analog stack adds histogram-based tone matching plus a physics-informed optical renderer for grain and per-channel halation, with those effects driven by parameters regressed from the reference image. On 350 self-supervised pairs, the color stage reaches 22.05 dB PSNR and 0.925 SSIM; the full pipeline lands at 21.72 dB / 0.923. The color path is reported at 5.2 ms for 1080p, or 192 FPS, and can export a standard .cube LUT for use in existing grading and editing tools.

For graphics programmers and technical artists, the practical takeaway is that open-set look transfer is getting closer to something shippable: fast enough for interactive use, portable enough for offline tools, and flexible enough to target a specific reference instead of a prebuilt look library. The paper also calls out a second collapse mode, residual-scale collapse, and argues that...