# Neural rendering experiments: scope and methods

These notes accompany “DLSS 6 could rewrite games while you play.” The title describes the author's proposed direction; these experiments demonstrate image processing, not generated game logic or an announced NVIDIA DLSS 6 feature. Measurements are a snapshot of the experiments through September 21, 2026 in America/Los_Angeles. The adjacent `evidence.json` contains selected values and SHA-256 hashes of the original local reports. It is a curated ledger, not a complete reproducibility package; the private NVIDIA model/runtime is not distributed.

## The experiments are separate

1. **NieR in-process Streamline bridge.** Windows D3D11 game, second D3D12 device on the same RTX 5090, shared GPU resources/fences, pinned signed NVIDIA runtime 310.8.0.0 and NR adapter 2.13.0.0. The shown paired readback uses RGBA8 (DXGI format 28). Earlier bridge tests also exercised RGB10A2. Synthetic zero motion, candidate real D24S8 depth, fallback camera, HUD in source. The adapter drops the requested per-frame history reset. The live functional result does not establish temporal quality or a performance uplift.
2. **Synthetic diagnostic harness.** Known static 1280×720 room/spheres with analytic depth and valid zero motion. Identity-copy check precedes inference. A changed nonblank image establishes execution. Attenuation tests create a fresh feature before every pass to remove retained history. Direct-NGX reset tests instead keep one feature alive, change its history through feedback, then restore source input and intervene on native reset.
3. **Memory tests.** DXGI process-local LOCAL-segment usage includes the bridge, feature resources, pools and caches. Feature destruction and physical resource retirement are different observations. The eight-cycle standalone chart uses three 4K viewports per cycle; the five live-game disable cycles are a separate experiment. The old 6,197 MiB observation included a previously retained allocation and is not the size of five fresh models. Full shutdown returned near a stable baseline in the measured cycles, not a proof of indefinitely zero growth.
4. **Ghostwire external NVIDIA presenter.** Windows Graphics Capture and DirectComposition in a separate process, direct NGX with locally reconstructed NR parameters. It uses native resets, synthetic guides and final color including UI. It never feeds its enhanced overlay back as the source. Four live F9 enable events and five NR ON transitions were observed without initialization errors after the lifecycle fix. Forza profiles/configuration exist, but gameplay is not validated. This is not reuse of NieR's injected DLL or a native engine integration.
5. **Model reconstruction.** Static block/layout analysis and tensor round-trip fidelity are distinct from numerical execution.153 packed records decoded into 851 tensors with 2,394 bytes still uninterpreted. The parameter count 154M comes from NVIDIA's report, not an independent sum of recovered trainable weights. Native byte-exact validation covers the FFN-expansion operator only: 24 comparisons and 4,751,360 FP8 values across blocks 31/38. The complete forward and fine-tuning remain unfinished.
6. **Qwen stills and offline video.** Qwen-Image-2.1 is a separate model, not a checkpoint swapped into NVIDIA's runtime. The native 4K precision comparison uses one matched 40-step call per profile, fused code and tiled VAE, with a 3840×2176 generation cropped to 3840×2160. It is not an untiled reference or a real-time benchmark. The later Ghostwire run uses 2048×1152 masters, all 40 steps and full token processing, with NVFP4 and fusion. It is paused at 71/297 anchors. No MLP merging, temporal flow or original-frame compositing occurs in those masters. The planned 4K temporal reconstruction has not run for this photoreal job.

## Image and video handling

- The NieR slider uses 1920×1080 web derivatives of the paired before/after 4K captures. The inspector loads lossless 3840×2160 WebPs whose decoded RGB pixels match the original GPU readback BMPs exactly. These are two sides of the same captured frame.
- The five-pass artifact example, where included, is an actual NVIDIA screenshot. It is not a matched comparison with the one-pass hero. Style, pose, framing and capture time differ.
- Plot PNG/SVG files are generated with Matplotlib from recorded values. Bar baselines are zero. The memory plot also starts at zero; the 588 KiB post-shutdown change is stated numerically rather than exaggerated by cropping its axis.
- Qwen preview derivatives discard the near-opaque generated alpha and retain RGB color. Original RGBA PNGs and receipts are unchanged. Web previews are downsampled and compressed for delivery, with sizes disclosed; they should not be used for bytewise numerical comparison.
- The 8.5-second Qwen video is a 68-keyframe snapshot. Each frame is held until the next selected source timestamp. It contains no interpolated frames or audio. Its low update rate is separate from any single-frame editing defect.
- The paired Qwen sequence viewer uses all 71 completed keyframes and the corresponding source-working images from the same receipts. Source timestamps identify the original clip. The sources are downsampled from native 4K to the model's 2048×1152 working input before both sides are prepared for web display.
- Qwen inspection loads 2048×1152 WebP derivatives (quality 96), with no resizing. These preserve native pixel dimensions but are lossy visual-inspection copies. Numerical reference PNGs remain unchanged. Full-size files load when requested.
- Attractive stills are subjective visual evidence. We have not established temporal acceptance, a causal diagnosis of the rejected sequence, or a controlled benefit from revising the prompt.

## Public sources

- [NVIDIA's DLSS 5 research overview](https://research.nvidia.com/labs/adlr/DLSS 5/)
- [NVIDIA technical report, section 2.6](https://research.nvidia.com/labs/adlr/files/DLSS 5_Report.pdf)
- [NVIDIA product announcement and controls](https://www.nvidia.com/en-us/geforce/news/dlss-5-3d-guided-neural-rendering/)
- [Inspected Streamline 2.14.1 release](https://github.com/NVIDIA-RTX/Streamline/releases/tag/v2.14.1)
- [Streamline programming guide](https://github.com/NVIDIA-RTX/Streamline/blob/v2.14.1/docs/ProgrammingGuide.md)
- [Streamline manual-hooking guide](https://github.com/NVIDIA-RTX/Streamline/blob/v2.14.1/docs/ProgrammingGuideManualHooking.md)
- [Deferred-resource implementation](https://github.com/NVIDIA-RTX/Streamline/blob/v2.14.1/source/platforms/sl.chi/generic.cpp)
- [Qwen-Image-2.1 model card](https://huggingface.co/Qwen/Qwen-Image-2.1)
- [Pinned Qwen image-editing guidance](https://github.com/QwenLM/Qwen-Image-2.1/blob/fb7ae1d1f9611cd91524d03c53c5246b36ac8577/prompt_rewrite/prompts/system_prompt_edit.txt)

The public Streamline source is version 2.14.1, while the inspected binary adapter is 2.13.0.0. Source-level mechanisms are supported by observations from the pinned binary, not assumed to be an exact source match. The runtime audit covered named public packages; it did not inspect every authenticated NVIDIA developer download.