DLSS5 HYSTERIA/THOUGHS/NEWS (Patreon)

Hello. I’ll start from the beginning:DLSS 5 Hysteria.While everyone is cramming DLSS/NR into everything imaginable—games, movies, videos, and more, I tried to find a practical application for this new tech. I was surprised to discover that it isn't a typical AI model that redraws pixels by regenerating pixel tiles (the way generation usually works, by reassembling the image from small pixel clusters like 2x2, 4x4, or 8x8 blocks but with its own generated upon original tile). Instead, DLSS/neural rendering leaves the original pixels and the image's tile structure untouched 100%. It only adjusts the brightness and saturation of individual existing pixels by establishing relationships with their neighbors. This preserves the original pixels while recalculating lighting and surface structure—for instance, making metal look more realistic by adding effects that might be missing from a 3D material not set up by a professional, or improving the look of skin, as in my case. This is best illustrated by the character model I am currently working on:

Left is original, Right is with Neural Rendering(NR) applied. Because character is having very dark skin, the difference in this one is pretty noticeable. What it gives? It gives more correct micro specular, because it's trained on faces, and it makes skin look not so dry, but also not too wet. Nose tip, cheeks and lips look more wet, while forehead and other skin over face is not touched, which look more correct. Also contact areas between eyelids and eyes is look way more soft and correct, i trying to solve this issue for 2-3 years already, tweaking SSS and other parameters with masks and other shit. I made good progress, but it still too far from actual smooth SSS shading for eyes and eyelids, NR helping with it not only fixing brightness levels, but also giving me clear result that i need to achive for any specific separate chracter, which is super useful.Second thing that it gives it's screenspace lighting dependence, To get what i mean, look on lower line of the jaw, in original, because shader is not good enough - lighting got swallowed that reflected yellow light rays on the sharp border of the jaw, while NR draws it, taking in to account background's general color, walls color, which is soft yellow. And the last, but mostly attractive to me is Contact Ambient Occlussion, This example not that good for explaining it, but i will give simple example, but if taking in to example this one, just look on contact line between lips, in original shading it is kinda glowing, because imperfections of the shader's SSS, while NR processed gives correct soft darkening in the contact zone line between lips. But in other cases i checked some of my renders with girls wearing clothes, and it gives that correct soft darkening effect in the contact zone between clothing and body. And overall i came to conclusion:This is a very good automated post-processor; it doesn't redraw anything or replacing elements with its own versions, unlike 99% of AI models. Secondly, I didn't see anything here that couldn't be achieved manually with shading and manual tonemapping, for instance, skin sheen could be handled via additional shader settings (which I plan to create soon to achieve a more accurate result, now i have clear reference), though that would require a significant amount of time for painting B&W masks and running hundreds of tests. Most likely, a custom texture pack will need to be created for each character, because every face is different, skin tones is different and skin structures as well, but it's doable.Contact Ambient Occlussion(AO). Also doable, even more, i already implemented it and also using it as mask for SSS on the contact zones, to simulate blood outtage from that area to make squeezed zones look less transluscentive, but it also not perfect because of Blender's Cycles limitations of the renderer. Will try to make some adjustments to shader's AO system as well, but to be honest NR doing it super good and in realtime. Okay i just made this gif with one of recent renders with Sonya to show what i mean exactly:

As you can see it adds correct smooth shading, that cycles right now can not provide in such deep level, because of it's software limitations of ray bouncing calculating and etc, well technically i can turn everything on maximum, rays length, ray bounces, caustics and etc, but one render in 1080p will take coule of days(weeks) to render it on 4090, so kinda questionable, but NR gives that little touches just like that, in realtime. Fucking cyberpunk POGBut as well it have it's bad sides, at least one i noticed very much, it's that one with Tanya, i don't know why, but only with Tanya it works weird, with any other characters with dark skin it works well, with Susan, with Jade, Calypso, but with Tanya it works super strange, NR just trying to remove her eye shadows too much and making her face look super wrong lol, just look at it:

As i said this issue only with Tanya, no matter in what scene she is, in dark one or in very bright one, it always tries to eliminate her eye shadows makeup lol.So I’ve concluded that this technology is an excellent post-processing method for a finished piece; it doesn’t alter the composition, color balance, or levels(if you not overplaying with it's settings ofc), but simply corrects the renderer's imperfections by filling in missing shadows, lighting, and reflection data on surfaces. Most importantly, it leaves the image's original pixel structure intact. I used to use AI occasionally to fix poor polygonal hair models, regenerated cut segments of image with hair to merge it back after re-processing, just to make them look a bit better so the whole piece held together. You don't want everything looking perfect only to have hair that looks like it’s from 2004 lol, completely ruining the viewer's overall impression. Hair is a key detail the brain focuses on when building associations and verifying the reality of what we see; if a specific feature—like the eyes, hair, nose, or lips—looks jarringly out of place or poorly animated, the brain quickly loses interest in the subject in that case. Someday I’ll graduate to full-blown hair simulation, but that shit will probably require a quantum computer—not just to make it look good, but because it would stretch the animation workflow for a single static scene to a full month lol, render times will increase as well. But that’s a problem for the future.NEWS:In 26 of this month i gotta go to the border with Armenia to make a Visa-run, to renew my period of stay in Georgia, yeah i need to do that every year, every 365 days) so i hope nothing wrong will happen and i'll get back the same day to continue doing my nasty things. That's all news for today)