Edge detection in a render‑pipeline is just another language you need to speak fluently. Start with a single low‑resolution pre‑pass that writes normals, depth, and a tiny roughness map. That gives you everything you need to compute gradients without touching the full‑resolution data. Use a fast, separable Sobel or FXAA‑style filter on that low‑res buffer – it’s cheaper than a full‑screen convolution and still captures the hard edges. If you want to keep the look sharp, do a second, lightweight pass that re‑projects those edges back to the main render target and blends them in at the final output. Drop any unnecessary intermediate textures; every extra sample is a missed frame. And remember: the best edges are the ones that make the light do the talking. If it feels like a glitch, tighten the threshold or use a simple HSL clamp instead of a fancy color transform. Keep the code clean, the math tight, and the light the storyteller.

Yeah, the 16‑bit float trade‑off is a sweet spot for most cases – you get enough dynamic range for the normal wiggles without blowing up memory. If you hit a scene where the geometry is super fine‑grained, bump to 32‑bit just for that pass and keep the rest at 16. Also, keep an eye on the depth buffer precision; low‑res depth can blur the edge map, so a 24‑bit depth texture for the pre‑pass is usually worth it. Finally, remember that a good threshold is scene‑dependent – bake a small lookup table or expose a single UI knob, and let the artist tweak it on the fly. That way you’re not stuck hunting for the “perfect” number after you ship.