I'm always amused by naive implementations like this.

Like, the idea is to save bandwidth & storage, but because you don't compare the file sizes, the smaller file gets discarded?

So now every time that happens, you're storing & serving possibly multiple times as many bytes because of lazy code.

LMAO

I remember being so confused by AeroWings at first because "fighter jets but no fight???" and also "why can't I pull harder?"

... and now I own every game in the series that got released in the West and I actually understand and quite enjoy the physics.

A strange hypothetical, I know, but I've been thinking: I love planes - Wouldn't it be awesome to own a private jet? Maybe even one with gold fixtures!

It'd cost so much to operate though... That's "multiple homes" money. Not to mention the environmental impact! I could never.

A screenshot depicting the main menu of the 240p Test Suite for NES, ported by Damian Yerrick. The "Sharpness" option has an arrow next to it. This screenshot was taken with a real NES connected to a Koryuu transcoder over composite, connected to a RetroTINK-4K via component cables. The sample rate per line was set to a multiple of 341 (240p). The text is very sharp, but has subtle colour splotches around each letter. In addition, the screen has subtle diagonal lines across the green/brown background, and there are sawtooth-like artifacts on the character graphic and text box's border, a consequence of the fixed-function comb filter used by the Koryuu transcoder. Text: 240p Test Suite Menu PLUGE Gradient color bars SMPTE color bars Color bars on gray Color bleed Monoscope Convergence Gray ramp Solid color screen IRE Sharpness Chroma crosstalk Overscan Safe areas CPU clock speed ←1/2→ | ↑↓A: Select | NTSC

A screenshot depicting the main menu of the 240p Test Suite for NES, ported by Damian Yerrick. The "Sharpness" option has an arrow next to it. This screenshot was taken with a real NES connected to a RetroTINK-4K via composite cable. The text is sharp, but has a slight horizontal blur; As a consequence, it has more visible fringing, but less expansive "rainbow" artifacts are visible. The green background is clean, and both the character's lines and text box border are more solid, with only a subtle ripple effect thanks to better comb filtering separating luma (brightness) from chroma (colour). Text: 240p Test Suite Menu PLUGE Gradient color bars SMPTE color bars Color bars on gray Color bleed Monoscope Convergence Gray ramp Solid color screen IRE Sharpness Chroma crosstalk Overscan Safe areas CPU clock speed ←1/2→ | ↑↓A: Select | NTSC

I had the dumb idea to hook up a Koryuu composite-to-component transcoder to the RetroTINK-4K to see what it'd look like.

While the RT4K has more robust filtering, the result of being able to dial in the sample rate over component is even composite video pixels get really sharp!

I don't think I'll miss this era of the internet. Then again, I don't know what's next. We've gone from "hey look at this neat thing" to being crowded out by grifters, ads, and the worst, most moronic flamewars of the dial-up era, amplified, and instead of getting moderated, they're rewarded for it.

It sucks because being trans isn't a "certain people" thing, though I guess I can be happy that most people never feel the need to think about their gender identity in the first place. At least, not in a way that makes them think anything more than "must try harder to [gender]".

I first saw this when I came out as trans in 2023. Because it was a Big Deal™ by then, *certain people* saw it as a wave (or narrative arc) and took bets on "who's next".

In reality, I've known since 2010, but my mental health took a nosedive ignoring it and I finally had access to care here.

People see "patterns" and "connect the dots", even if there isn't actually pattern there.

I've realized the internet is basically seen as reality TV; Anyone of note is just a character. It's "stereotypical housewives gossiping about soap operas" behaviour for the modern age.

It's possible that the demo you played used DLS and that either deemed too unstable/not broadly supported enough or just not pursued further, or they attempted to do a software synthesizer of their own and found it took too many resources. Or you have an AWE card and the demo defaulted to using it.

The PC release shipped with .SF2 banks for use with AWE cards with the PSX samples - or close to it, One Winged Angel was fully rendered. It has 4 different bundles with MIDI tracks intended for different devices - AWE, general MIDI, Yamaha XG, and "YGM", which I'm not sure what corresponds to.

The frame-to-frame I'm less worried about, because it's being fed motion vectors and should (god I hope) compensate. What I want to know is what happens when you do a 360. Will skintones / makeup change constantly?

There are a lot of fears about what this means for gaming; Art direction has been a huge part of gaming since day zero, because the kind of game we're playing is a *video* game. If things look different from moment to moment, or the incentive is for devs to just let the AI handle it... What then?

Yeah, when TVs cranked the brightness, turned on frame generation, upped the sharpness and contrast and saturation, all to look like the brighter, prettier set in a sea of other TVs. It looks good to the old monkey/lizard brain in the moment... But on closer inspection, it really doesn't.

I remember when I played Battlefield Bad Company 2 and basically being snowblinded on maps with snow or sand because of all the bloom and fullbright. I'm glad we moved on from that for a while but like, half of this can be accomplished mostly for free with ReShade.

An image from 2008 called "Real Super Mario" by Pixeloo. It depicts a hyperrealistic version of Mario's head, with individual pores, wrinkles, eye veins, ear folds, stitching, stubble, and generally way too much detail. It is extreme in its uncanny valley effect.

I don't really have a conclusion to this because between now and when it launches presumably there will be a lot changed, but on aggregate... Man I really don't get Digital Foundry's coverage at all.

Shout out to 18 years ago.

Source: www.deviantart.com/pixeloo/art/...

A demo screenshot from Resident Evil Requiem with "DLSS 5 OFF", showing the overpass supports. through the gaps, road signs can be seen further into the distance, with localized lighting.

A demo screenshot from Resident Evil Requiem with "DLSS 5 ON", showing the overpass supports. Through the gaps, the road signs contrast less and "fade" into the overall contrast level of the supports, reducing the depth and rejecting the objects' local illumination.

Worse still, this makes background elements that are partially obscured appear to be taken as having the same depth as the objects in front. There are details behind this support that are further away, but are flattened to the same relative contrast as the support itself.

A demo screenshot from Resident Evil Requiem with "DLSS 5 OFF", showing the streetlight behind Grace. It is visibly brighter than the rest of the scene, suggesting that it is on and emitting light.

A demo screenshot from Resident Evil Requiem with "DLSS 5 ON", showing the streetlight behind Grace. It is no brighter than the rest of the scene, suggesting that it is either off and not emitting light, or broken/dim.

A demo screenshot from Resident Evil Requiem with "DLSS 5 OFF", showing the reflection of a car's brake lights reflecting off the fender of the car behind. The reflection is bright and contrasts against the background. The red colour is casting/splashing around the area, tinting both the rest of the fender and the rear of the car ahead. The pillar/lamppost next to the reflection has a gold/bronze tint, presumably from the light above.

A demo screenshot from Resident Evil Requiem with "DLSS 5 ON", showing the reflection of a car's brake lights reflecting off the fender of the car behind. The reflection is no brighter than the scene around it, and the colour cast of the red brake light is largely removed. The pillar/lamppost next to the reflection is largely pure black.

In addition, emissive lights within the scene are muted substantially, in an apparent attempt to normalize the image. This is something that can probably be tweaked, but it's shocking that this was the first demo they showed given how much is wrong with it.

A demo screenshot from Resident Evil Requiem with "DLSS 5 OFF", showing a sign in a store window for what appear to be a pack of cigarettes on a purple background. Beneath, the sign reads "THE REAL DEAL" against a yellow background.

A demo screenshot from Resident Evil Requiem with "DLSS 5 ON", showing a sign in a store window for what appear to be a pack of cigarettes on a purple background. Beneath, the sign reads "PNZ REAL DEAL" against a yellow background. The image is sharpened, and texture is added to the purple background.

A demo screenshot from Resident Evil Requiem with "DLSS 5 OFF", showing the building behind Grace. The textures suggest concrete and stonework, with visible grain.

A demo screenshot from Resident Evil Requiem with "DLSS 5 ON", showing the building behind Grace. The texture has been smoothed out and has lost definition. The texture is smoothed more like marble, and the brickwork is barely legible. Interestingly, this is one case where contrast appears lower than the original.

However, lighting and facial detail is only part of the story. Background elements also get touched, and this is often unfortunate. In the first Resident Evil screenshot with Grace, store signs are less legible, and background textures have less definition.

A comparison shot from Nvidia's DLSS5 website, showing the "Zorah" demo in a daylit scene. A seam divides the middle of the female character between "DLSS 5 OFF" on the left and "DLSS 5 ON" on the right; The left has a warm, muted colour grade, with subtle facial shadows. The right increases contrast and sharpness, while adding light and grading the scene in cooler tones. The character's face appears much shinier on the right, with visible lipstick and eyeshadow, with prominent eyelids.

A demo screenshot of Resident Evil character Leon S. Kennedy from Resident Evil Requiem, with "DLSS 5 OFF". Leon has a determined look on his face in a dark scene with subtle lighting. He has short facial stubble, and his eyes glint with what is presumably the sole light source. Shadows are soft and subtle, implying the light source is weak or far away, implying a moonlit scene or a streetlight some distance away. The direction of specular light reflections and details mostly or entirely in shadow strongly suggests light coming from only one direction.

A demo screenshot of Resident Evil character Leon S. Kennedy from Resident Evil Requiem, with "DLSS 5 ON". Leon has what now looks like a confused look on his face in a dim scene with stark lighting. He has longer facial hair, and his eyes glint with multiple light sources, implying studio lights or a row of lights such as a warehouse. Shadows are harder and sharper, implying the light source is strong or close by. The areas newly-highlighted by specular reflections (lower eyelid, under-chin hair, visible ear) imply light is coming from the front, below, and to the sides. Leon's chin has been given additional shape from the shadow, giving him a more chiseled look with a more prominent cleft chin.

In the daytime shots Nvidia provides, it looks somewhat natural in direct light, save the added makeup & removal of colour grading. In dark scenes, extra light comes from nowhere. Tellingly, sharp shadows are drawn where none existed, implying a very bright light source nearby.

A still from one of the shots used in one of Emily Young's videos, showing her under a desk and checking the framing of the shot. The light is set up to bounce off the floor, which lights her face mostly naturally in this dark scene, but her arm and clothing in the direct path of the spotlight are brightly lit and shining, showing the difference between the direct output of the light and the intended effect.

I'm not good at it, but lighting is one of the most fundamental parts of photography and especially film. Every shot is lit with purpose so the camera captures the image as intended - not necessarily realistically - which usually means the face is well-lit, even in dark scenes.

A selfie of Emily Young, taken in 2024. The sun is to her back, backlighting her hair and leaving her face in indirect light. The shadows of her face are soft and indistinct due to being lit by bouncing light rather than direct light.

A still from a video by Emily Young, showing her in an environment with high-powered lights. Shadows are more distinct and her skin has specular highlights that reflect the lights illuminating her. One of the lights is visible in her glasses; A softbox with honeycomb to diffuse the light and control the amount of spill, specifically used to reduce this effect. However, it is very difficult to eliminate in such intense direct light.

A still from a video by Emily Young, showing her in a more open environment with high-powered lights further away and diffused by both the soft box and bouncing from the ceiling. Shadows are slightly less distinct, but her skin still has some specular highlights that reflect the lights illuminating her. The lighting conditions are many times brighter than the room is normally.

Skin isn't that shiny normally, and there isn't enough light in that scene for there to be that much visible red in the ears. What we're seeing is a model trained on professional photo/video shoots where subjects are wearing makeup (yes, men too) and/or blasted by studio lights.

I feel like I need to explain some things.

Clear facial distortion aside, some people like the DLSS 5 image because it's high-contrast and looks like film looks like. But that's not what reality looks like. Cameras need a whole lot of light to not look like a fuzzy mess.

Way to just trigger a core memory out of nowhere. I still have mine...

There are definitely ways ML could be used to guide the rendering process that don't involve just throwing out the render result and pasting genAI onto it.

*This* is basically like saying you can have your own fine wine at home if you leave your grape juice behind the fridge.

I wasn't sure if DLSS5 was just doing img2img generation before I sat down and looked at it on something other than my phone, and this basically slapped me in the face with the answer.

A screenshot from Nvidia's DLSS5 announcement, showing a character from Starfield with "DLSS ON". The face appears to be replaced by a diffusion-generated one, with differently shaped and coloured eyes (the right eye, viewer left, is substantially larger than the other), and apparently incorrect perspective. Next to the "DLSS OFF" shot, it looks as though the face has been lasso selected in Photoshop and rotated/moved. The chin appears broader than the original.

A still from Nvidia's DLSS5 announcement demo reel, showing the "DLSS OFF" version of the same character. The face is generally symmetrical.

"Character rendering is transformed."

I legitimately thought about ending the thread with this but I'm waiting for someone else to do it because it would be hilarious

As sad as it may be, as it is now, it's not & can't be your friend; It doesn't feel nor think, nor has intent.

Like any program, it follows rules: It is - reliably - both reliable & unreliable in precisely the same way. Without RNG, it's a book - the same every time you open it.

It's all based on probability, tuned by training data & the parameters given at runtime, plus some good old RNG. Prompts are categorized by meaning, then the most probable word is chosen, word-by-word, based on each previous word until the most likely output is reached.