One run, two ways to query the score
The schedule moves on. The state may take another route.
Both paths start from the same noisy point and keep guidance in every update. The noise-conditioned score receives a label that decreases throughout the run. The noise-unconditioned score selects its implicit noise state from the position it actually reaches.
The gray and green arcs belong to the same connected manifold. The green arc satisfies the constraint; the remaining gray arc is also valid data. Stronger guidance pulls toward the green region. Larger steps make larger updates and use fewer sampling steps over the same run. Adjust either slider while paused to compare the same stage of sampling.
One sampling timeline. Every displayed waypoint is an actual update.
Noise-conditioned path
The scheduled σ keeps decreasingThe blue training band shrinks with the prescribed schedule. A guidance-driven displacement can leave it before the next score query.
Noise-unconditioned path
Implicit σ follows the actual positionThe selected band responds to the point’s position. The same guidance term continues steering toward the green region.
The dots mark actual sampling updates: coarser steps produce fewer dots and larger jumps. Motion between dots only helps playback; it is not another solver step. Blue gradually turns red as the illustrative score error grows outside the scheduled training band. The band ratio is distance to the manifold divided by the scheduled band's half-width; a ratio above 1 marks an off-band state. Training bands are schematic.
One clock, different noise states
Scheduled versus state-selected noiseThe blue schedule reaches σ min at the final update. No terminal noise value is imposed on the green curve.
What the illustration shows. Without guidance both constructions can return to the manifold. Gentle guidance can steer both into the constrained region. Stronger guidance or coarser updates can move a state away from its scheduled training band; the red portion illustrates possible accumulated learned-score error. The state-only score adapts to the current position while guidance continues. These are mechanism illustrations, not measured neural-network trajectories.
Which scales contribute at the current position?
The position selects relevant noise scales
The sampling clocks
Both paths use the same initial state, guidance coefficient λ, and update budget K over the same time interval. Increasing α reduces K and enlarges each time step. The guidance coefficient λ remains unchanged.
The noise-conditioned prior follows a probability-flow Euler step tied to the decrease in variance. Without guidance, its update is Δx = ½(σk2 − σk+12) s(xk, σk). Its motion and shrinking training band therefore advance on the same noise clock.
The state-only field uses position-dependent scaling and receives no scheduled noise label. Its implicit scale follows the current position; it need not wait for the blue schedule to reach its final value.
Guided updates retain both the prior and task gradient, with positive diagonal damping. Outside the scheduled training band, a smooth illustrative error field varies with the position and assigned label. It depicts possible learned-score error, not measured network behavior or an undefined analytic Gaussian score. Playback speed changes viewing time only.