lonewolf0622/HelixSR

Drop-in FSR 3.1 upscaler for Direct3D 12 that runs DLSS's network (Model E) on AMD GPUs

What it solves

HelixSR allows AMD GPU users (RDNA 1 and newer) to use NVIDIA's DLSS Model E neural upscaler. It provides a way to get DLSS-like image quality and scaling modes—from native (DLAA) to Ultra Performance—on hardware that does not have NVIDIA Tensor Cores, CUDA, or NVIDIA drivers.

How it works

Instead of creating a new model, HelixSR runs NVIDIA's existing DLSS network as standard Direct3D 12 compute shaders. The setup process downloads the official DLSS DLL from NVIDIA, builds the network specifically for the user's GPU, and integrates it into games via three methods:

  1. OptiScaler: Uses OptiScaler to route DLSS inputs to HelixSR.
  2. FSR Drop-in: Replaces the game's FSR 3.1 DLL with HelixSR.
  3. Experimental Loader: Uses a custom loader (dxgi.dll) to trick the game into thinking an NVIDIA GPU is present, running HelixSR as nvngx.dll.

Who it’s for

Gamers using AMD RDNA 1 or newer GPUs on Windows or Linux (Proton) who want better image quality in Direct3D 12 games than what is provided by their native upscalers.

Highlights

  • Hardware Agnostic: Runs without Tensor Cores, CUDA, or ROCm.
  • Faithful Reproduction: Mimics DLSS behavior, including camera jitter and render sizes, without adding default sharpening.
  • Flexible Integration: Supports multiple installation paths including a drop-in replacement for FSR 3.1.
  • Customizable Sharpening: Optional support for NVIDIA or RCAS sharpening filters via a configuration file.

Related

  • Project
    nanoporetech/dorado

    A high-performance analysis engine for Oxford Nanopore reads that provides GPU-accelerated basecalling, alignment, and error correction for genomic data.

  • Project
    wildminder/AI-windows-whl

    A curated collection of pre-compiled Python wheels for AI/ML libraries that are difficult to install on Windows, eliminating the need to build from source.

  • Project
    notatestuser/ds4-control

    DS4 Control is a macOS menu bar application that manages and monitors the local DeepSeek V4 inference server, providing resource tracking and automated model downloads for local agentic coding.

  • Project
    spiritbuun/buun-llama-cpp

    buun‑llama‑cpp is an experimental fork of llama.cpp that introduces Variable‑Bit‑Rate (VBR) KV‑cache quantisation, a family of low‑bit “turbo” codecs (including trellis‑coded variants), native safetensors/EXL3 loading, and utilities for multimodal and MoE models. VBR dynamically compresses the attention cache layer‑by‑layer, letting users maximise context length on a given GPU while keeping quality high. The project is aimed at researchers and developers who already run LLM inference and need to push memory limits.