Frame Generation Gets a Referee, Not Just Another Contender

Every few months, the PC graphics world discovers that a category it thought was settled is not settled at all. Frame generation was supposed to be NVIDIA’s thing — DLSS-G, locked to RTX 40 and 50 cards, with AMD’s FSR 3 and Intel’s XeSS playing catch-up. Along comes Apex Compute with Apex Neural Rendering, or ANR: a frame generation implementation in beta that claims to run on NVIDIA GeForce RTX 20 through 50 series (plus GTX 16), AMD Radeon RX 5000 through 9000 (plus the Steam Deck), and Intel Arc A and B series, on Windows 10 and 11 and on Linux and SteamOS through Proton. That is, roughly, every GPU anyone has bought in the last seven years.

That breadth alone explains the attention. But if you go to the repository expecting to read a neural network, you will find something stranger and, arguably, more useful: a measurement apparatus.

What is actually in the repo

An honest inventory first, because the README invites a specific misreading. ANR has three planned components: ANR FG (frame generation, in beta now), ANR SR (super resolution, “beta soon”), and ANR RR (ray reconstruction — denoising for ray tracing, also “beta soon”). The source code for the ANR models themselves is “coming soon.” What ships today is a set of release packages, installers, and — the part whose source is actually in this repository — a tool called nr-record.

So the open-source project, at present, is not the neural network. It is the instrumentation around it. That is not a dismissal. The instrumentation is the part with lasting value, and it is the part nobody else has built properly.

The FG beta itself is distributed as two packages: a full one for games that don’t already use OptiScaler, and an add-on for games that do. The distinction tells you the architecture. ANR FG is not a standalone middleware an engine integrates — it is a drop-in replacement for the frame generation DLL a game already loads, delivered through the OptiScaler ecosystem.

Riding on OptiScaler

The mechanism deserves a plain-language explanation, because it is both clever and slightly legally fraught, and the README is candid about it. OptiScaler is a GPL-3.0 licensed shim that intercepts a game’s calls to vendor upscaling and frame generation libraries and redirects them to alternatives. ANR FG installs itself as an OptiScaler “custom” extension: the game still shows the NVIDIA frame generation toggle in its settings menu, still believes it is talking to NVIDIA’s nvngx_dlssg interface, and instead gets Apex’s implementation. The DLL is even named to match. The README notes that OptiScaler’s GPL-3.0 source is attached to each release — the compliance box, ticked.

This is the same pattern the modding community has used for years to bring DLSS-class features to games that never shipped them, or to AMD and Intel cards that NVIDIA’s SDKs officially ignore. What Apex adds to the pattern is a competitor with its own model, distributed through the community’s existing plumbing. The game-side experience is deliberately boring: you flip on the NVIDIA frame generation option, and ANR generates the frames instead.

Boring is the point. The entire value proposition rests on the swap being invisible to the game.

nr-record: the referee

Here is where the repository earns its keep. Comparing frame generators is methodologically miserable. The usual approach — run the game twice, record the screen, squint — is confounded by everything: different input sequences, different camera paths, nondeterministic engine timing, capture-card compression. You are never sure the two generators saw the same problem.

nr-record solves this by separating capture from generation. It records, losslessly, the exact inputs a game hands to its frame generation — the rendered frames, motion vectors, depth, whatever the interface carries — while the game runs normally. During recording, the game displays only real frames, so the generated ones can’t contaminate what’s on screen. A captured frame costs roughly 26 MB at 3840×2160 and about 6 MB at 1806×972, which tells you these are uncompressed or near-uncompressed buffers, not video. The tool then replays a recording through any frame generation DLL you have — ANR FG, or the game’s own DLSS-G — and produces videos and images from identical inputs. A bundled viewer plays the results in sync, side by side, with a divider, pixel zoom, and markers for close inspection.

Crucially, nr-record “includes no frame generation.” It is a neutral party. You can use it to compare ANR against DLSS-G without trusting Apex’s marketing, or to compare DLSS-G against itself across driver versions, or — with the upscaler recording mode, which captures super resolution inputs and replays them through DLSS, FSR, or XeSS — to do the same for upscaling. The replay tooling runs on Linux under Proton as well, with a single command path that handles both generators and assembles the comparison. There is even a documented on-disk format, which suggests Apex expects others to build tooling against these recordings.

This is the kind of infrastructure that usually stays locked inside a vendor’s QA lab. NVIDIA can compare DLSS versions against each other because it has the harness; the public gets cherry-picked screenshots. nr-record hands the harness to anyone. If ANR the product vanished tomorrow, nr-record would still be a contribution — arguably the more durable one, because it outlives any particular generator it judges.

One wrinkle: replaying DLSS-G recordings requires an NVIDIA RTX 40 or 50 series GPU, since you need NVIDIA’s DLL and its hardware gates to run it. The referee is neutral; the participants are not.

What “neural rendering” means here

The name invites confusion, and the confusion is instructive. In the research sense, neural rendering descends from NeRF — neural radiance fields, the UC Berkeley work that reconstructs a 3D scene as a continuous function queried along camera rays, synthesizing photorealistic novel views from a sparse set of photographs. AWS’s explainer frames NeRFs as networks that learn scene geometry and render views from angles never captured; the broader neural rendering literature describes a family of techniques that replace hand-built rendering rules with learned ones — denoising, inferring lighting, generating unseen perspectives.

NVIDIA has spent the last year claiming the term for its own stack. Its RTX Neural Rendering announcement covers RTX Neural Shaders — small networks embedded in programmable shaders — plus neural texture compression (up to 7× VRAM savings), neural materials (up to 5× faster processing), and a neural radiance cache that learns multi-bounce indirect lighting. That is neural rendering in the literal sense: networks participating in the synthesis of pixels.

Frame generation is something adjacent. It does not render the scene; it interpolates between frames the game already rendered, using motion vectors and depth to guess what the in-between frame looks like. Calling it “neural rendering” is marketing-adjacent — accurate in that a network is involved and pixels come out, generous in that the network is filling gaps rather than illuminating geometry. Apex is hardly the worst offender here; the whole industry has been stretching the label. But readers who arrive expecting NeRF-class scene reconstruction will find a very fast, very specialized interpolator instead.

(An aside for anyone about to search for documentation: “ANR” is a crowded acronym. The provided research surfaced, among other things, an opioid dependency treatment clinic, a glossary entry for organs authorized but not recovered, and a towering figure of Telugu cinema. Search accordingly.)

The honest limits

Start with the obvious: this is a beta, the super resolution and ray reconstruction components are not yet available, and the model source is promised rather than delivered. Until the source lands, the “open-source project” is a distribution and measurement framework around a closed binary. That is a legitimate choice — training and shipping a cross-vendor frame generator is expensive, and Apex presumably wants to keep the model itself proprietary — but it should be named for what it is.

Second, the README publishes no benchmarks. There is a test kit for users to run and file performance results as issues, which is a crowdsourced-evaluation strategy: let the community generate the numbers, with nr-record providing the apples-to-apples evidence. Refreshingly honest, and also a signal that Apex knows its claims need third-party verification. What the actual quality looks like — artifact behavior in disocclusions, UI elements, fast motion — the sources simply do not say. Anyone telling you ANR FG matches or beats DLSS-G right now is ahead of the evidence.

Third, frame generation’s known weak spot is latency, not throughput. NVIDIA’s own announcement pairs DLSS 4’s multi-frame generation with Reflex 2’s Frame Warp, which it says cuts latency by up to 75% compared to native rendering — an acknowledgment that generating extra frames adds input lag unless something actively counteracts it. The ANR documentation says nothing about a latency countermeasure. That gap may be handled and undocumented, or unhandled; the sources don’t distinguish.

Finally, there is the structural fragility. ANR FG works by impersonating NVIDIA’s DLL interface inside a GPL shim, on GPUs NVIDIA’s SDK was never meant to support, across DirectX 12 and Vulkan, Windows and Proton. Every one of those layers moves. The OptiScaler project absorbs the brunt of the maintenance, and Apex’s add-on package — for games already running OptiScaler — is effectively an admission that the community built the road and ANR is driving a truck on it.

Why it matters anyway

Strip away the branding and the situation is this: a third party has entered the frame generation market with a cross-vendor implementation, distributed through open community infrastructure, and — this is the unusual part — shipped a rigorous, open comparison tool alongside it that can judge its own product against the incumbent on identical inputs. Vendors do not typically arm the referees.

If ANR FG is good, the GPU-vendor lock-in that defines the upscaling era erodes a little further: a Steam Deck owner and an RTX 50 owner running the same frame generator is a genuinely new arrangement. If ANR FG is mediocre, nr-record still changes the conversation, because for the first time the public can verify vendor frame generation claims instead of taking marketing screenshots on faith. Either outcome disciplines the market. The beta’s quality is an open question; the tooling’s value is not.