From external feedback (2026-05-08): a 'Zheng et al. CVPR 2025' decoupled G-buffer paper apparently argues for separate color vs G-buffer encoders that can be pruned independently. Underlying insight is sound — depth/MV/normals carry less information than RGB, so smaller capacity + aggressive quant on the G-buffer branch is plausibly free quality.
Verification needed first: I could not locate the specific paper from the citation provided. Confirm the actual paper (title, arxiv ID, conference proceeding URL) before designing against it. Possible the citation is misremembered or the paper is named differently.
If verified, candidate v6.2.1 modification:
- Current concat-fusion (T1, oss/sr/v6/concat_fusion.py) treats color + G-buffer uniformly through one 1×1 → depthwise → 1×1 stack
- Proposed: small dedicated G-buffer pre-encoder (e.g., 6 channels → 4 channels via 1×1, INT4 quantized) → concat with color features → existing fusion stack
- Net effect: G-buffer branch gets aggressive quantization without affecting color quality; total params lower with same expressivity
Acceptance gate: +0.1 dB PSNR or -10% inference latency at iso-quality.
Companion: docs/architecture/2026-05-08-v62-arch-v4-spec.md
From external feedback (2026-05-08): a 'Zheng et al. CVPR 2025' decoupled G-buffer paper apparently argues for separate color vs G-buffer encoders that can be pruned independently. Underlying insight is sound — depth/MV/normals carry less information than RGB, so smaller capacity + aggressive quant on the G-buffer branch is plausibly free quality.
Verification needed first: I could not locate the specific paper from the citation provided. Confirm the actual paper (title, arxiv ID, conference proceeding URL) before designing against it. Possible the citation is misremembered or the paper is named differently.
If verified, candidate v6.2.1 modification:
Acceptance gate: +0.1 dB PSNR or -10% inference latency at iso-quality.
Companion: docs/architecture/2026-05-08-v62-arch-v4-spec.md