Crow - DLSS Rendering Tool

Current public source release: v0.7.3-alpha1 Windows x64 · NVIDIA RTX · Experimental / alpha software

Crow is an experimental Windows rendering/conversion tool that combines NVIDIA DLSS Neural Rendering, Frame Generation / Multi Frame Generation, NVIDIA Optical Flow, external render guidance, Auto Depth, and an optional SEA-RAFT neural optical-flow backend in one workflow.

This GitHub package is source-only. NVIDIA runtime DLLs, the NVIDIA Optical Flow SDK, PyTorch, SEA-RAFT source/weights, FFmpeg binaries, downloaded model weights, and local virtual environments are intentionally not vendored.

Feature overview

• DLSS Neural Rendering for images and video, including configurable image iterations.
• Recursive DLSS5 image passes: DLSS5 Passes can run 1–8 complete passes, feeding each completed image back into the next pass independently of Iterations.
• Frame Generation / Multi Frame Generation (FG/MFG) with capability-driven output multipliers and an optional NVAPI model preset path.
• NVIDIA Optical Flow (NVOF) with full-resolution and multi-scale motion paths, predictive reconstruction, confidence, visibility and uncertainty handling.
• SEA-RAFT neural optical flow with Spring-S / Spring-M models, selectable resolution/refinement controls and a CUDA PyTorch runtime.
• External Motion / Depth guidance from EXR sequences, including calibration and shared NR/FG guidance paths.
• Auto Depth using a pinned Depth Anything V2 Small ONNX model through ONNX Runtime DirectML.
• Video processing and portable packaging through FFmpeg/ffprobe, with diagnostics, parameter persistence and motion preview tooling.

v0.7.3-alpha1 highlights

Recursive DLSS5 image re-render

Image Conversion adds DLSS5 Passes with values 1–8 (default 1). Each completed DLSS5 image output becomes the input to the next pass.

DLSS5 Passes is separate from the existing per-pass Iterations option. For example, 3 passes × 4 iterations performs three complete image re-render passes, each using four internal evaluations.

SEA-RAFT Runtime Hotfix5

The current source includes the restored, validated SEA-RAFT runtime fixes from the Hotfix5 lineage:

• avoids the PowerShell $Args automatic-variable collision;
• validates .venv\Scripts\python.exe immediately after venv creation;
• validates CUDA PyTorch and torchvision independently;
• provides mainland-China mirror-first installation with official fallback where appropriate;
• uses GitCode mirror-first checkout for SEA-RAFT with official GitHub fallback;
• uses hf-mirror first for model prefetch in CN mode, then official Hugging Face fallback;
• loads Hugging Face checkpoints with RAFT.from_pretrained(..., strict=False);
• supports --check-only Spring-S / Spring-M model-load smoke tests.

Existing rendering pipeline retained

The release keeps the established V0.7.2 alpha6/sr2 pipeline:

• unified NR, FG/MFG, and NR → FG/MFG video conversion;
• NVOF full + multi-scale motion paths with predictive/temporal reconstruction;
• optional SEA-RAFT tunable neural motion backend;
• External EXR Motion / Depth guidance;
• Auto Depth;
• capability-driven FG/MFG multiplier handling;
• optional NVAPI FG preset integration;
• persisted image/video parameters and motion preview tooling.

Inherited architecture lineage

The current release retains the earlier Crow motion/rendering work rather than replacing it. The inherited lineage includes Adaptive Stable Motion, Adaptive Reliable Motion, Temporal Consensus Motion, Spatial-Temporal Dual-Path Motion, Visibility-Aware Uncertainty Motion, and Predictive Multi-Scale Motion Reconstruction. These names are preserved in the historical changelogs and contract tests.

Parameter persistence still uses per-control reset behavior. There is deliberately no Reset All action.

Build

The only root build entry is:

BUILD.bat

Build Center options include:

1. Unified NR + FG full build — CN sources
2. Unified NR + FG full build — Official/global sources
3. FG diagnostic standalone — CN sources
4. FG diagnostic standalone — Official/global sources
5. Portable build — CN sources
6. Portable build — Official/global sources
7. Refresh NVIDIA NVAPI SDK

See docs/BUILD_zh-CN.md for the detailed Chinese build guide.

Required / external components

A normal full build may require or acquire the following outside this repository:

• Visual Studio 2022 C++ Build Tools and Windows SDK;
• NVIDIA display driver;
• NVIDIA Optical Flow SDK 5.x;
• NVIDIA DLSS / NGX SDK checkout;
• user-supplied nvngx_dlssnr.dll appropriate for the installed GPU generation;
• DLSS-G runtime;
• TinyEXR;
• FFmpeg;
• Python and Auto Depth dependencies;
• CUDA PyTorch, SEA-RAFT source, and Spring-S / Spring-M weights when SEA-RAFT is installed.

The build scripts keep these downloaded/runtime components outside the tracked Git source through .gitignore.

Output layout

A completed build uses the following structure:

dist/
├─ Crow-DLSS-Rendering-Tool.exe
├─ Crow-DLSS-Rendering-Tool-Image.exe
├─ runtime/
│  ├─ nvngx_dlssnr.dll
│  └─ nvngx_dlssg.dll
├─ tools/
├─ video/
├─ auto_depth/
├─ sea_raft/
└─ models/

SEA-RAFT dependency boundary

SEA-RAFT is an independent Princeton Vision & Learning Lab project licensed separately under BSD-3-Clause. Crow ships an integration adapter and setup/runtime worker, but this source release does not vendor the upstream SEA-RAFT repository, PyTorch, or model weights. See sea_raft/README_zh-CN.md and docs/NOTICE.md.

Validation status

The v0.7.3-alpha1 source package passed the repository Python/contract suite used for this release (48 passed) and Python syntax validation of the SEA-RAFT worker. This remains alpha software; GPU/runtime behavior depends on the installed NVIDIA driver, runtime DLLs, SDKs, and hardware.

Release history

Detailed historical design notes, tests, and hotfix documents are kept under docs/. Current release-specific notes:

• docs/CHANGELOG_V0.7.3_ALPHA1_STRICT_IMAGE_ONLY_REWRITE.md
• docs/HOTFIX_V0.7.3_ALPHA1_SEA_RAFT_RUNTIME_HOTFIX5.md
• RELEASE_NOTES_V0.7.3_ALPHA1.md

License

Crow source is distributed under the GNU General Public License v3.0. Third-party components retain their own licenses. See LICENSE and docs/NOTICE.md.