DLSS 5 in Doom and Blood: Why AI ReShade Retro Remasters Go Wrong

In a recent hands-on experiment published by XDA Developers, writer Samarveer Singh took an RTX 4070 Ti, a copy of ReShade, and DLSS 5's DLL files, and pointed the resulting neural-rendering pipeline at a trio of 1990s classics: Doom, Doom II, and Blood (including Blood: Fresh Supply). The results were a fascinating, sometimes unsettling collision between modern generative AI rendering and 2.5D pixel-art history. While walls and corridors suddenly gained dramatic lighting and intricate surface textures, other parts of the experience—especially the grimy, oppressive mood of Blood and its pixelated enemy sprites—were mutated into something the original artists never envisioned.

To understand why slapping DLSS 5 onto a DOSBox window produces everything from impressive corridor makeovers to bizarrely muscular AI-hallucinated cultists, we have to look at both how these retro experiments are jury-rigged together and how far removed they are from NVIDIA's actual engineering blueprint.

How Do You Put DLSS 5 Into a DOS Game?​

If you know anything about modern upscalers and frame generators, your first question is probably the right one: Wait—how do you feed motion vectors from a 1993 MS-DOS engine into an RTX neural renderer?

The short answer is: you don't.

Games from the DOS era—whether they run on id Software's Doom engine or Ken Silverman's Build engine—were written long before DirectX 12, Vulkan, or modern shader pipelines existed. They do not output the scene depth buffers, velocity maps, or engine-level motion vectors that modern DLSS implementations expect. Even popular third-party injection wrappers like OptiScaler, which modders frequently use to swap or force upscalers in modern 3D titles, cannot simply hook into a classic DOS executable and find the necessary geometry data.

As Singh explains in his XDA Developers report, getting around that roadblock required treating each classic title as an individual compatibility puzzle. Instead of engine-level injection through OptiScaler, he used ReShade to load DLSS 5's DLL files—specifically the neural rendering library that modders have been circulating and injecting across unsupported games —and apply a real-time neural rendering overlay onto the game output. Because DOSBox itself can be finicky with full-screen overlays like ReShade, getting the pipeline running smoothly was half the battle.

This isn't an isolated science experiment, either. As Tom's Hardware reported on October 14, 2026, Ukrainian experimenter Valery Kryzhanovsky used a similar post-process setup—running the 1996 software-rendered version of Quake in DOSBox and hooking ReShade via RenoDX—to run a leaked DLSS 5 build in a 4x multipass configuration over the emulator's output frame. In both cases, the AI isn't collaborating with the game engine. It is essentially staring at a finished 2D picture of a retro frame and guessing what a "real" version of that scene ought to look like.

What NVIDIA Built DLSS 5 to Do—And Why Post-Process Hacks Break the Rules​

To appreciate why this distinction matters so much, it helps to look at how NVIDIA officially architected DLSS 5.

According to NVIDIA's Applied Deep Learning Research (ADLR) paper published in September 2026 and Phillip Singh's September 22, 2026 post on the NVIDIA Technical Blog, DLSS 5 is the company's first DLSS technology designed to generate final displayed appearance rather than merely reconstruct a higher-resolution frame. NVIDIA calls this approach 3D-Guided Neural Rendering.

In a genuine, developer-integrated DLSS 5 pipeline—which NVIDIA designed to run locally on GeForce RTX 50 Series GPUs at up to 4K resolution—the model is strictly tethered to the game engine:

• Engine Grounding: It uses a one-step pixel-space diffusion model conditioned on the current rendered frame, artist-authored geometry, lighting buffers, carried temporal state, and game-engine motion vectors.
• Deterministic Stability: It operates on a strict one-frame-in, one-frame-out model so that textures and lighting don't boil or morph as the camera moves.
• Developer Art Direction: Studios get granular controls, including Structure Intensity and Tone Intensity sliders, model selection across scenes, semantic AI masking, and engine-level masks. For example, in NBA 2K27, developer Visual Concepts used per-pixel uplift control masks and tone controls so DLSS 5 could enhance skin subsurface scattering, contact shadows, and light transmission through hair without distorting the players' scanned facial geometry.

Now compare that tightly controlled, engine-guided architecture with a ReShade overlay sitting on top of DOSBox.

When you strip away the motion vectors, the geometry buffers, and the semantic masks that tell the AI "leave this character's face alone, only shade the marble floor," the neural renderer has to improvise. And when an AI improvises over 320x200 pixel art, things get weird fast.

Case Study 1: Doom and Doom II Get a Fresh Coat of Industrial Paint​

In XDA Developers' testing, Doom and Doom II came out looking the best of the bunch—at least as far as the environments were concerned.

In the original id Software classics, chunky low-resolution textures and flat sector-based lighting are part of the visual DNA. When Singh enabled DLSS 5, the industrial corridors of Mars and Earth immediately picked up convincing surface relief. Metallic wall panels, ledges, and stonework looked sharper and easier to tell apart.

The standout change, however, was the lighting. Using DLSS 5's neural lighting—stacked across three neural rendering passes to make the effect pop—transformed the mood of the levels. Yet the experiment also exposed a weird schism right down the middle of the screen:

• The Environments Changed: Walls, floors, and architectural surfaces absorbed the neural passes, looking almost like a modern remake.
• The 2D Sprites Mostly Resisted: Doors, pickups, and 2D enemy sprites remained largely unchanged compared to the overhauled backgrounds. Singh noted that while Doom's demons picked up slightly more defined facial features and abdominal muscles, the 2D sprites mostly sat inside a completely altered 3D-looking environment.
• Atmosphere Shifts: In Doom II, while stonework and wall panels looked cleaner and more defined, the neural pass also brightened the presentation considerably, changing the brooding atmosphere of the original levels.

Still, for Doom and Doom II, the trade-off felt like an intriguing curiosity—a fresh coat of paint that made thirty-year-old corridors pop without completely ruining the experience.

Case Study 2: Blood Suffers an Aesthetic Identity Crisis​

If Doom showed the flattering side of neural rendering on retro geometry, Monolith Productions' 1997 cult shooter Blood (tested alongside its modern port Blood: Fresh Supply) demonstrated how quickly generative rendering can trample a game's soul.

Anyone who played Blood back in the late nineties remembers its unmistakable vibe. It is a game built on grime, gothic dread, and pitch-black humor—flickering torches, blood-streaked stone crypts, and oppressive, shadowy interiors that make up roughly 80% of the campaign.

According to Singh's account, turning DLSS 5 loose on Blood made him want to switch the feature off immediately:

1. The Grime Was Scrubbed Away: DLSS 5 gave the stone walls pronounced texture and depth that the 1997 Build engine could never render natively, and it added dramatic lighting to the crypts. The problem? That brighter, cleaner lighting completely bleached out the filthy, claustrophobic dread that makes Blood work.
2. Hallucinated Enemy Sprites: Unlike Doom, where the 2D sprites barely budged, DLSS 5 latched onto Blood's pixelated enemies and rewrote them into what Singh described as "disturbingly detailed human figures" with pronounced clothing folds and musculature.

Think about what is happening technically here. Those muscles and fabric creases aren't hidden inside the original 1997 sprite sheets waiting to be "upscaled." Because DLSS 5 is a generative model trained on real-world visual appearance priors, it looks at a cluster of brown and grey pixels shaped like a cultist and synthesizes a brand-new, high-detail human body on top of it. It is not restoring the original art; it is replacing it with an AI interpretation.

Why "More Detail" Does Not Equal a True Remaster​

What makes these experiments so valuable for PC gamers and graphics enthusiasts isn't that anyone should actually play Blood or Quake through a 3x or 4x ReShade neural pass. In fact, Tom's Hardware pointed out that running 1996 Quake through a 4x multipass DLSS 5 filter in DOSBox looked noticeably worse—and overly darkened—compared to proper engine-level ray-tracing mods like Quake RTX or NVIDIA's own Quake II RTX. Furthermore, as MakeUseOf noted in its own October 2026 testing of classic Doom with DLSS 5 neural rendering, forcing heavy neural passes onto a game that could otherwise run on a toaster introduces a real performance and responsiveness penalty on modern hardware.

Instead, these tests highlight a critical truth about retro graphics and modern AI:

• Low Resolution Was a Canvas, Not Just a Limitation: In the 1990s, artists didn't just shrink photorealistic assets down to 256 colors. They drew sprites and textures specifically to read clearly at low resolutions, relying on stark contrast, exaggerated silhouettes, and Moody palette swaps—and your imagination—to fill in the blanks.
• Generative AI Hates a Vacuum: When you feed sparse, pixelated 2.5D frames into a generative neural renderer without engine masks or developer guardrails, the model fills every blank space with literal, high-frequency surface detail. A dark, ominous smear of pixels that felt terrifying in 1997 gets turned into a brightly lit stone wall and a bodybuilder in a robe.
• Engine Integration Matters: The gap between NBA 2K27's targeted, masked DLSS 5 implementation and a brute-force ReShade overlay proves NVIDIA's own thesis: generative rendering needs engine-level guidance and artist controls if it is going to respect the original art direction.

What If You Actually Want to Remaster Your DOS Classics?​

If reading about AI-mutated Barons of Hell has you itching to replay your DOS backlog on Windows 11, you don't need a leaked neural rendering DLL or a triple-pass ReShade stack to get a gorgeous experience. Veteran PC gamers generally stick to methods that respect the original rendering intent:

1. Use Modern Source Ports: For Doom, Doom II, Quake, and Blood, community source ports and dedicated remasters (such as GZDoom, Ironwail, or Nightdive Studios' KEX Engine releases) let you run native 4K resolutions, high frame rates, and optional Vulkan/DirectX renderers while keeping the original sprite art crisp.
2. Try Engine-Level Path Tracing Mods: If you want modern lighting in classic shooters, look for mods built directly into the engine—such as Quake II RTX or Sultim Tsyrendashiev's ray-traced classic Quake and Doom mods—where light sources actually understand the 3D level geometry.
3. Use Clean CRT Shaders Instead of Generative Filters: Ironically, if you use ReShade or a modern DOSBox fork like DOSBox Staging, the best visual upgrade for pixel art isn't AI generation—it's a good CRT scanline and phosphor shader. Retro sprites were drawn for curved glass monitors, and restoring that display characteristic often makes 1990s art look far more authentic and detailed than an AI hallucination ever could.

Experiments like pushing DLSS 5 into DOSBox are part of what makes the PC platform great. We tinker, we break renderers, and we push silicon to do things the engineers never intended. Still, as Singh rightly concluded after his tour through Doom and Blood, technical wizardry is no substitute for artistic intent. Sometimes, an ugly, pixelated 1993 demon is already doing its job just fine.