Give each sampler
a reason to exist.

Why this scene

The sharp lens concentrates emitted paths onto a small receiver region. Merging can collect nearby light vertices there; connection remains useful for less constrained transport elsewhere.

Try it

Compare connection only → merging only → both. Then reduce the merge radius. Look at the caustic and diffuse sphere separately, including noise, smoothing and cost.

The local VCM implementation weights connection/merging using a nearby-photon density proxy and applies a fixed 0.5 factor to their sum. These are experimental weights, not full path-density MIS. A pleasing combined image does not establish the published VCM estimator's consistency or efficiency. This scene exposes that distinction.

Independent draws repeatedly forget where the bright paths were. Metropolis sampling can spend additional work near a successful path after finding it. The source is deliberately finite: this camera-path implementation cannot invent a missing delta-light sampling strategy.

Try it

Toggle acceptance/chain diagnostics, then compare large-step probability 0.3 versus 1.0. Count bootstrap time and mutations, not only displayed frames. The all-large-step setting still uses Metropolis weighting; it is an independent-proposal MCMC baseline, not ordinary PT.

The implementation records and perturbs primary random numbers. It should not be described as the complete original path-space mutation algorithm.

Lazy primary samples support variable-length random-number streams without eagerly modifying every coordinate. They are a different implementation of mutation sampling, not a promise of a different transport capability.

Try it

Compare large-step probabilities 0.1, 0.3 and 0.6, restarting each run. Small steps refine discoveries; large steps explore elsewhere. Check both colours: high acceptance while one region stays dark is a chain-mixing problem.

Load this same scene in recorded-vector MLT for the implementation comparison. Different images and startup budgets cannot establish that one variant is intrinsically faster.

Local material sampling knows the BSDF, but not which directions lead to illumination elsewhere. A learned directional distribution adds that information.

Try it

Enable Show field after the training passes. Compare guiding strength 0 versus 0.5, resetting between runs. Inspect where the arrows concentrate, and include training cost before judging performance.

Audit finding: the guided branch uses a mixture denominator, but the cosine branch does not evaluate that same mixture. The sampling distribution also needs a solid-angle audit. Current images demonstrate learning behavior; they cannot support an unbiased variance-reduction claim. PDF repair is a prerequisite to a fair benchmark.

True manifold methods move a path while enforcing reflection/refraction constraints. That can address specular transport which arbitrary vertex connections almost never satisfy.

What works here

The current mode records specular chains and perturbs the primary samples more gently at specular bounces. Enable chain and mutation visualization to inspect accepted proposals.

There is no Newton constraint solve in this implementation. The Newton controls are inactive, and recorded Newton steps remain zero. A new scene cannot supply the missing algorithm: this is a useful fixture for developing it, not a showcase of proven manifold acceleration.