hi sota

Prompt

Image Denoising SOTA — Current Ideas & Evidence **Exported:** 2026-09-06 · **Status:** idea-discovery → benchmark-structure mining (compute-free, CUDA blocked) ## End goal A state-of-the-art real-noise image denoiser (SIDD sRGB ≈40.8 dB ceiling, DND hidden test) built on a **structural, DeepSeek-MLA-class insight** — not an incremental +0.1 dB trick. SOTA is a measured-benchmark claim, never a reviewer-opinion claim. ## The thesis (one line) **The noise — not the image — is the low-rank latent.** Real sensor noise is generated by a tiny parametric law (`σ² = a·y + σ_read²` per camera + low-dim ISP), so it can be *compressed into its latent and sampled from*, collapsing the data bottleneck that actually caps real-noise denoising. This is the denoising analog of MLA's "the KV object is low-rank — compress it, and the win is free." ## Measured evidence (execution-grounded, CPU-only) | Gate | Result | |---|---| | G1 noise channel low-rank (`σ²=a·y+b`) | **PASS — R² = 0.970** on 3,200 pairs | | Heteroscedasticity | **144× variance range** (std 0.013 dark → 0.158 bright) | | Noise realization structure | white / full-rank (patch-PCA top-4 = 7.6% ≈ white-noise 6.3%) | | Synthesis fidelity (held-out 640) | **near-PASS**: power law `σ²=0.0262·y^1.61` → held-out dev 6.4% max / 2.1% mean (vs 76% linear) | **Measured noise law:** `σ² = 0.0262 · y^1.61` — exponent 1.61 > 1 is the **super-Poisson / ISP-gamma signature** (train R² = 0.9992). The linear form was the wrong functional model; the power law is the corrected synthesis recipe. ## The SOTA-candidate (DeepSeek/Kimi transplant) 1. **Data-axis exponent change:** sample unlimited calibrated noise from the fitted law over any clean corpus (DIV2K / BSD / OpenImages) → collapse the ~160-scene SIDD bottleneck (the axis "more parameters" cannot fix). 2. **Sparse noise-routed MoE** (DeepSeekMoE / Kimi K2): fine-grained experts + one shared "universal denoiser" expert; **aux-loss-free bias balancing** `b_i += γ·sign(imbalance)` — because patch statistics skew pathologically toward flat midtones. 3. **Low-rank latent conditioning** (MLA analog): condition on a compact σ̂-map (pair-supervised blind estimator) instead of making the network re-learn the channel. 4. **Multi-target supervision** (MTP analog): predict clean + residual-noise + downsampled-clean; drop auxiliary heads at inference. 5. **Muon / MuonClip optimizer** (Kimi K2): token-efficient, spike-free — underused in restoration. **Falsifier (preregistered):** a "cheating oracle fed BOTH realizations" must reach ≥41.3 dB, else the information ceiling is real and no architecture moves it. ## Idea pool (ranked, with status) | # | Idea | Status | |---|---|---| | 1 | Noise-latent synthesis (data-axis lever) | **active — measured: σ²=0.0262·y^1.61, held-out ≤6.4% dev** | | 2 | Sparse σ-routed MoE + shared expert + bias balancing | candidate — needs GPU to falsify | | 3 | Cross-realization consistency (T3) | survived hostile review; +0.05–0.10 dB honest margin | | 4 | Target-noise-aware loss (`v_c = v_y/N`) | from GT frame-average finding; small but principled | | 5 | L2-over-L1 (median-vs-mean skew from clipping/gamma) | cheap; needs SIDD to confirm | | ✗ | semantic memory, predictive coding, multi-hypothesis WTA, trajectory teacher, CVaR, blind-spot | loss-decoration / published / wrong-objective | ## Next actions (aligned: measure → prove lever → design → train → benchmark) 1. ~~Finish noise-law model selection~~ → done: `σ² = 0.0262 · y^1.61` (super-Poisson, ISP/gamma signature). 2. ~~Validate corrected synthesis on held-out~~ → done: power law holds at 6.4% max / 2.1% mean dev. 3. Re-fit the law on SIDD-Medium sRGB when downloaded (form is sensor-universal; parameters are not). 4. Design + preregister the architecture A/B (GPU-gated). 5. Restore CUDA (nvidia-smi permission failure on RTX 5070 Ti Laptop) → reproduce baseline → train. ## Constraints & sources - CUDA unavailable (`torch.cuda.is_available()==False`); all current work is CPU-only analysis. - Legacy set: grayscale, 3,200 pairs, 2×-SR confound (256 clean / 128 noisy) at `C:\Users\Vikaash\Documents\HA1(SEM)`. - Artifacts: `NOISE_LATENT_FINDINGS.md`, `denoising_pilot/noise_latent_*.json`. --- ## Appendix A — Complete idea history (all discussed) **Prior ideas (earlier Codex session, HA1/SEM restoration):** EdgeRestore compact residual CNN (29.09 dB incumbent) · wider 5×5 local operator · null-space/detail forcing (rejected) · denoise-then-SR serial (rejected, −1.37 dB) · D4 test-time aug (+0.03 dB) · spectral augmentation (closes 57% of measured spectral gap; never wired into training) · degradation conditioning · evidence-guided fast weights · trajectory/flow teacher (deferred) · grow/prune NAS (deferred) · LoRA policy search (deferred) · semantic/detail memory (deferred). **Round 2 — 5 personalities × 6 ideas (30 total), clustered by underlying angle:** - *Posterior/objective surgery:* learned VST front-end · noise-realization gain injection · multi-expectile composition · CVaR worst-decile loss · codebook classification decode · WTA multi-head decode · heteroscedastic moment-matching · SNR-gated Laplacian-band supervision. - *Variance reduction (ensemble/soup/consistency):* twin cross-realization distillation · zero-cost ensemble bake-in · jackknife greedy soup · orthogonalized momentum (Muon) · INT8 QAT · realization-dropout · burst-determinism consistency · saccade probe-and-fuse · oracle estimator fusion · cross-scale posterior consistency. - *Noise/physics + sensor conditioning:* VST · noise-realization gain · divisive normalization · retinal adaptation · fingerprint-leakage penalty · restoration-rendering decoupling. - *Structure/detail prior injection:* evidence-gated predictive coding · center-surround DoG bank · sigma-delta cascade · structure-tensor directional prior. - *Data/split hygiene:* scene-disjoint folds · burst-determinism · fingerprint-leakage · rendering decoupling. - *FLOP reallocation:* RGGB re-mosaic · quadtree tokenizer · boosting cascade. **Round 3 — 4 reasoning-only agents (DeepSeek engineer, physicist, forensic, renegade), converged on:** (1) GT is a noisy ISP-rendered frame-average → `v_c = v_y/N` ≈ 15–35% of SOTA residual; (2) the pair `x1−x2` is a free, texture-free exact σ-map; (3) a real PSNR delta comes only from information-injection or capacity/scaling-curve change — not loss decoration, not "forbid the conditional mean" (Blau–Michaeli). **SOTA-candidate (DeepSeek/Kimi transplant):** data-axis synthesis · sparse σ-routed MoE + shared expert + aux-loss-free bias balancing · low-rank latent conditioning (MLA analog) · multi-target supervision (MTP analog) · Muon/MuonClip optimizer. **Discovery candidates (discovery-first reframe):** - D1: noise-variance exponent is set by the ISP mapping (`p = 2 − 1/q`), ISO-invariant — measured `p = 1.61`. - D2: the SIDD leaderboard ceiling is partly GT-target-noise artifact (quantified `v_c = v_y/N`). ## Appendix B — Hidden connections found | Pattern | Connection | |---|---| | DeepSeek MLA (KV is low-rank) | ↔ the noise channel is low-rank (`σ² = a·y^p`, R²=0.97) — "the noise, not the image, is the latent" | | Measured exponent 1.61 | ↔ super-Poisson = ISP/gamma signature, not sensor shot noise | | Dark-decile mismatch (linear law) | ↔ read-noise floor + sRGB linear toe + tone curve — predicted by D1 | | Pair `x1 − x2` | ↔ free exact σ-map → routability needs a *learned* σ̂, not a naive local statistic (G2 r=0.46) | | SIDD GT = frame-average | ↔ `v_c = v_y/N` → leaderboard ceiling partly benchmark artifact | | DeepSeekMoE / Kimi sparsity | ↔ sparse noise-routed MoE (expert per noise regime) | | MTP / MuonClip | ↔ multi-target supervision / token-efficient spike-free training | | AutoSOTA / BFTS / QD / MAAR / ScientistOne | ↔ execution-grounding / tree search / diversity archive / dynamic routing + isolation / evidence-chaining | ## Appendix C — Novelty-gate verdict (HONEST, 2026-09-06) **The core "noise-latent synthesis" idea is OCCUPIED** — it is an active subfield, not a discovery: - **YeTI** (arXiv 2607.09193, 2026-07) — "You Only Need Two Noisy Images for Real-World sRGB Noise Generation": per-camera variance–intensity profiles; synthesizes realistic sRGB noise from two noisy images. - **NAFlow** (2022, normalizing-flow sRGB noise) · **NeCA** (neighbor-correlation-aware) · **C2N** (unpaired image-to-image noise) · **Pseudo-ISP** (2024, learn pseudo-ISP + raw noise) · **DualDn** (2024, dual-domain differentiable ISP). - "ISP distorts noise; sRGB noise deviates from Poisson-Gaussian" is textbook (noise-level-function / NLF literature, 2015+). - **SIDD itself ships the noise level function (NLF) in per-DNG metadata** — the noise-vs-intensity curve is a dataset-provided quantity. **Still possibly unclaimed (needs deeper novelty check before any claim):** - D1's closed-form "exponent `p = 2 − 1/q`, ISP-set, ISO-invariant" as an explicit testable law. - D2's quantified leaderboard GT-noise floor (AIM 2025 discusses 200-vs-500-frame GT quality only anecdotally). **Consequence:** no discovery is established yet. The measured results confirm a known law on one dataset. The discovery-first discipline worked: it stopped the loop from claiming as new what YeTI/NAFlow/etc. already own.