From "I direct the art" → "I speak the machine's own language, and I edit with a scalpel."
How machines represent, process and generate images, video and sound — taught on real Tom Horn assets, with dashboards you can actually turn the dials on.
Right now, when you want a change, you describe it. "Make the cherry read better on the dark reel." "Same style, different fruit." "Cut it out cleanly." And then you iterate — three rounds, four rounds — because description is lossy.
"Make the edges cleaner." becomes:
"The matte is hard-thresholded. Re-export with partial alpha, choke 1 px if the measured silhouette calls for it, and remove the white-matte contamination from the edge RGB; that's why the fringe only shows on the dark reel."
Both sentences describe the same wish. The second one cannot be misunderstood — not by an artist, not by a tool, not by a model. That's the whole track.
Phase I — SEE · what an image is → Phase II — OPERATE · surgical editing → Phase III — THE MODEL'S EYE · how machines see & generate → Phase IV — TIME & SOUND · video, audio, capstone
1 of 16 modules published. Sequenced work-first: the things that land on Samo's desk on a Tuesday come first; the model theory comes once you can name what the model is manipulating.
A provably surgical edit · a Style Parameter Sheet v1.0 the studio keeps · three new family symbols that pass a blind test.
Zoom to 64× until the cherry dissolves into labelled hex cells. Kill channels, starve the bit depth, drop the resolution. Live histogram, LAB readout, alpha breakdown.
Aha: drop to 3 bits and the highlight bands — nothing was blurred, the machine just ran out of numbers.
Threshold, choke, feather and break the premultiply. Live out = src×α + bg×(1−α) per pixel, plus a matte quality report with a verdict badge.
Aha: flip the backdrop white → black with a contaminated fringe. Every "dirty cut-out" ticket in THG history, explained in one gesture.
Type nine numbers into a 3×3 grid, get every filter in Photoshop. Live sliding-window arithmetic and a luminance cross-section that shows sharpening overshoot as a literal spike.
One picked colour plotted live in RGB / HSV / LAB / CMYK at once, with the whole symbol recolouring as you drag. Out-of-gamut pixels flash.
Draggable curves, a deforming 3D LUT cube, and a .cube export. Grade one symbol, snap the whole family into consistency.
A nine-axis radar chart per image. Two symbols that "obviously match" show a 40% divergence on outline weight and noise floor — and that's exactly what made the new one feel off.
Original / reconstructed / amplified-difference. Watch the WILD lettering melt while the cherry survives. The whole "AI can't do text" question, answered in one image.
An actual denoising loop running in canvas. Steps, CFG, sampler, seed. At CFG 20 you watch the burn happen.
Before/after, ΔE-LAB difference, changed-pixel count, pass/fail against an "only this region may change" contract. This is how you prove surgical.
Resample & Scale · Compression Forensics · Mask & Blend Workbench · Transform & Outpaint · Style Dial Board · Embedding Neighbourhood · Control Signal Extractor · Denoise Ladder · Frames & Easing · GOP & Bitrate · Waveform & Spectrogram. Full specs in CURRICULUM.md §2.