How an idea
becomes light.

One prompt. Two generators.

The prompt and seed enter two independent paths. A pinned SDXL Base model makes the digital teacher image with 30 denoising steps. Separately, frozen SDXL text encoders make text features for the optical student. The student also gets Gaussian noise determined by the seed. It does not receive the teacher image or run SDXL denoising at inference.

Writing phase, not pixels.

A shared, prompt-conditioned planner organizes the scene on a 32×32 feature grid. The latent-lift writer combines those features with seeded noise and expands them into detailed, 1000×1000 phase commands. Red, green, and blue are written as separate sequential exposures. SLM1 carries the learned phase pattern; SLM2 is held flat in this checkpoint.

The optical simulator propagates each phase field through a model of the bench. That model includes the measured 49×49 wavefront correction. Its three channel outputs are combined into the optical simulation image.

The camera has the final word.

The deployment script sends the generated phase commands to the physical SLM setup and captures the red, green, and blue exposures one after another. It saves the raw camera frames. A fixed registration and processing recipe combines them into a 640×640 RGB image, including a one-pixel Gaussian blur and the recorded contrast, gamma, and brightness calibration.

That result is labeled physical bench in the gallery. If the computer or hardware is offline, the website remains visible; new capture requests wait until the computer reconnects.

What the scores really say.

NCC is normalized cross-correlation measured on full 640×640 RGB images. A score of 1 means the compared image patterns match closely; near 0 means little pixel-level agreement. The gallery shows simulation ↔ teacher and bench ↔ teacher. The saved comparison also reports bench ↔ simulation, which measures how closely the hardware follows the optical model.

A good bench-to-simulation score does not imply a good prompt-to-image result. In one arbitrary fox-prompt check with the 1,000-image checkpoint, simulation-to-teacher RGB NCC was 0.002 and bench-to-teacher was −0.072, although per-channel bench-to-simulation scores were about 0.68–0.70. That test showed the bench followed its simulation while the student had not learned that new scene. Results for other prompts can differ.