overview

Section F of the Feature Backlog: implicit diffusion: calibration and posterior samplers; likelihoods, polynomial and algebraic factors. Open rows only; a row that is done moves to Backlog Done 2026-11 with its id. Status, priority, ids and the column needs are explained in Feature Backlog, which also holds the dependency graph.

Sources: original to this repository; the sources of each row are in its source column.

idfeaturestatusneedsprioritysourcewhy
F1Graded block levels in ImplicitDiffusion: a fixed noise level per block that rises with the block’s age (old blocks nearly clean, recent ones noisier, free ones annealed), generalising today’s annealed-or-clean blocksproposed—lowLenticulumLM’s Incremental Next-Token Prediction (an idea, 2026-11-05)the schedule of Rolling Diffusion and Diffusion Forcing as a query option; holding clamps clean already helps (implicit.md §9); whether untrained-for schedules help is open
F2Calibration benchmarks for implicit relationspartial: circle, lemniscate, unbalanced rings (2026-11-05, lib/VariationalDiffusion.jl/examples/calibration.md); a pitchfork and a redundant arm remain—mediumboth reviews (LP-023)baseline measured: every off-domain query returns a root, the lemniscate’s recall is at best 0.52, the 10% ring is never found
F3Calibration remedies: off-relation detection (the full score’s norm, an energy threshold, a round trip), annealing the field levels, ridge residualspartial: round_trip and ridge_residual (VariationalDiffusion answers.jl), and the round trip, the ridge residual, coverage and annealing measured on the benchmark (2026-11-12 and 13, calibration.md “Remedies”, LP-030 Findings of the review of round_trip, LP-033 Findings of the review of the ridge residual); an energy threshold remainsF2mediumboth reviews; calibration.mdmeasured: both checks remove every off-domain root; coverage removes 9 of 11 and 96 of 99 spurious roots on the circle and the two circles with no loss of recall, but costs recall at the lemniscate’s crossing (0.52 to 0.31); the round trip also rejects true roots where the relation is parallel to the clamped coordinate (circle recall 0.86 to 0.74), which a ridge residual would not; annealing by continuation = (16, 4) lowered recall everywhere and left the lemniscate with no root; the ridge residual removes every off-domain root and loses no matched root at the low levels (spurious 11 to 3, 36 to 16, 99 to 27)
F4Diffusion posterior samplers, as query methods of ImplicitDiffusion beside root finding, RED-Diff (reddiff.jl), ProxDM (proxdm.jl) and replacement/RePaint (implicit_sample). Each is checked against the exact posterior of GaussianMixtureEps under a linear-Gaussian measurement, and compared on the calibration benchmark (rows below)proposedF2mediumDaniel, 2026-11-05 (“whatever useful sampling method appears”)today the library samples a query by replacement or optimises it (RED-Diff, ProxDM); a noisy or nonlinear measurement needs guided or particle samplers
F4a— DPS: the reverse SDE guided by through Tweedie’s estimate (arXiv:2209.14687; chung2023dps in refs.bib)proposedF2mediumas abovethe standard baseline; needs the network’s VJP, which the adjoints provide
F4b— ΠGDM: pseudoinverse guidance for linear measurements (Song, Vahdat, Mardani & Kautz, ICLR 2023; not on arXiv, metadata to check on OpenReview)proposedF2lowas abovecheaper and better-conditioned guidance for linear operators
F4c— DiffPIR as a sampler: half-quadratic splitting with the noise predictor’s Tweedie denoiser as the prior’s proximal step, then re-noising (arXiv:2305.08995; zhu2023diffpir)proposedF2lowas aboveprox_infer already does the splitting with a proximal network; this uses an ordinary noise predictor
F4d— DDRM: SVD-based sampling for linear inverse problems (arXiv:2201.11793)proposedF2lowas aboveexact handling of a linear operator’s singular directions
F4e— DAPS: decoupled noise annealing (arXiv:2407.01521)proposedF2lowas abovelarge moves early, against DPS’s local corrections
F4f— Twisted diffusion sampler: sequential Monte Carlo with twisting functions, asymptotically exact conditional sampling (arXiv:2306.17775)proposedF2mediumas abovethe one exact sampler of the family as particles grow: an oracle for the others, and weighted samples are a SampleBelief the graph can use
F4g— MCG: manifold-constrained gradient (arXiv:2206.00941)proposedF2lowas abovea correction to DPS-style guidance
F5Heteroscedastic and heavy-tailed likelihoodsproposed—lowChatGPT review §5
F6Polynomial factor: higher-degree ideal quotients, missing-data EMplanned—lowpolyfit.md
F7Algebraic factors: MLE existence by LP, Markov bases, splitting large tablesplanned—lowL-SPEC-004 Algebraic Factors Toric and Grassmannian open, §2.4
F8Polarity-dependent energies (allowed by the interface, used by no factor)proposed—lowL-SPEC-001 The Factor Interface open