Skip to content

Latest commit

 

History

13 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

RT4D Reconstruction Stack

Build

A custom Vulkan reconstruction pipeline for native GLB geometry on AMD GPUs. The current proof runs armored-sentinel-v1.glb through:

  1. GLB decoding and GPU vertex/index upload
  2. Base-color factor and texture ingestion
  3. Device-local texture upload through a staging buffer
  4. Per-material low-resolution G-buffer draws
  5. Temporal reprojection and confidence-weighted denoising
  6. Classical super-resolution, tone-map composition, and PNG readback
  7. Typed observability exports, motion sequences, and temporal metrics

This project is not NVIDIA DLSS. Its current super-resolution path is classical compute; ML super-resolution and frame generation are not proven.

Project status: proof-stage native GPU stack with verified material, motion, history-reset, and observability captures plus portable source-build CI. It is not yet a production renderer or an artist-approved material pipeline.

See the capability matrix for the exact boundary between verified, partial, and unavailable features, and the roadmap for the order in which the stack is intended to grow.

Verified proof

The verified v0.4 proof is a deterministic camera-orbit sequence with an explicit camera cut. It exports final composites, low-resolution internal images, per-frame temporal diagnostics, and full-sample plus post-warm-up GPU timing percentiles on an AMD Radeon RX 480 using RADV. Exact source, binary, shader, sequence, export, and capture hashes are in:

  • receipts/rt4d-reconstruction-stack-v0.4-receipt.json
  • receipts/armored-sentinel-motion-v0.4.run.log
  • receipts/armored-sentinel-motion-v0.4-metrics.json

Material status is provisional_source_materials: the reconstruction path now evaluates glTF base-color factors and base-color textures using five primitive draw ranges. The v0.3 texture is generated diagnostic fixture data embedded in the GLB; it is source-provided to the runtime but is not artist-authored or artist-reviewed look development. Metallic/roughness, alpha modes, emissive, normal maps, and full PBR remain outside this release.

See the observability contract for visualization and metric semantics, and the material policy for the source-UV and generated-planar boundary. The material classification remains provisional; v0.4 does not add artist review or full PBR.

The v0.4.1 maintenance proof also exercises every scene selector and hardens PNG publication. Taco, Battle, and Dragon now produce visible shader-path diagnostics using the Sentinel mesh as provisional geometry; they are not authored Taco, battle, or dragon assets. Captures are written to a .partial file, checked for a complete PNG envelope, and only then published. See receipts/rt4d-reconstruction-stack-v0.4.1-receipt.json and receipts/six-scene-v0.4.1/.

Build

Requirements: CMake 3.20+, a C++17 compiler, Vulkan development files, GLFW 3, glslc, and Python 3 when tests are enabled.

cmake -S . -B build-rt4d-recon
cmake --build build-rt4d-recon -j2

Run the proof

Run from the repository root so shader and fixture paths resolve correctly:

./build-rt4d-recon/mandala_rasterize recon \
  --asset=fixtures/armored-sentinel-textured-v0.3.glb \
  --frames=12 \
  --camera-sequence=orbit \
  --camera-cut-frame=6 \
  --sequence-dir=receipts/armored-sentinel-motion-v0.4-sequence \
  --observability-dir=receipts/armored-sentinel-observability-v0.4 \
  --metrics=receipts/armored-sentinel-motion-v0.4-metrics.json \
  --capture=receipts/armored-sentinel-motion-v0.4-final.png

The sequence uses frame-indexed camera positions rather than wall-clock time. Frame zero and the camera-cut frame explicitly reject and clear history. The observability directory receives cut and final keyframes; the metrics receipt contains camera samples, motion, confidence, residual, composite-delta, and GPU timing percentiles. --metrics requires --sequence-dir so composite deltas cannot silently become unobserved zeroes. Capture frame counts are bounded to 1–600.

For the factor-only comparison using the released Sentinel without an image:

./build-rt4d-recon/mandala_rasterize recon \
  --asset=armored-sentinel-v1.glb \
  --frames=3 \
  --capture=receipts/armored-sentinel-material-factors-v0.3.png \
  --debug=gpu-timer

Textured primitives without TEXCOORD_0 are rejected by default. The explicit diagnostic fallback is --missing-uv=generate-planar-labeled; generated UVs are counted and logged separately from source UVs.

The GLB fixtures are intentionally kept in this repository because they are small and form part of the reproducible proof contract.

The current window is intentionally fixed at 1280x720. Camera cuts use the same typed reset decision reserved for a future transactional resize, but swapchain, render-pass, pipeline, export, and history recreation is not yet implemented.

Agent Skill package

This repository includes a project-scoped Agent Skill at .agents/skills/rt4d-reconstruction-stack/. Install it directly with the standard Skills CLI:

npx skills add warheart1984-ctrl/rt4d-reconstruction-stack \
  --skill rt4d-reconstruction-stack \
  --agent codex

This is a source-repository install, not a listing in AMD's official skill catalog. The current AMD catalog uses SKILL.md packages rather than a root skill.yaml, and catalog federation currently requires an AMD-owned source repository. See the AMD Skills and Hyperloom notes for the verified interface, packaging boundary, and future integration seam.

License

No project-wide license is granted at this stage. Public visibility does not grant permission to copy, modify, or redistribute the project. Vendored third-party files retain the license terms and notices contained in those files.

About

Custom Vulkan reconstruction stack with reproducible GLB GPU capture evidence

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages