Watch

Meta Vens 2e54a4d380 Add pre-tonemap HDR bloom pyramid

2026-05-09 00:40:43 +01:00

11 KiB

Raw Blame History

TSR-Inspired TAA Spec

Aquarium needs an owned temporal resolver, not a copied implementation. The target is inspired by public TSR behavior: stable subpixel detail, aggressive history validation, visible debug terms, and special handling for pixels whose appearance is not fully described by motion vectors.

Inputs

The resolver owns these signals:

current HDR scene color
current ray travel
previous HDR history color
previous ray travel
current camera/Grid constants
previous camera/Grid constants
current and previous projection jitter
stochastic coverage from diegetic Grid surfaces

Later gates add:

field id
normal
explicit velocity
reactive mask
transparency/composition mask
volumetric accumulation terms
resurrection history

Current Pass Contract

The scene pass renders into an HDR texture:

rgb = linear scene color
a   = ray travel in world units

The pass does not tonemap. Tonemapping belongs after temporal resolve.

Reprojection

For the first implementation, motion is camera-derived:

Reconstruct the current world hit from current camera, current jittered ray, and current travel.
Project that world hit into the previous camera basis.
Remove previous jitter to find the previous history UV.
Sample previous history if the UV is valid.
Compare previous history travel against the reprojected point's expected distance from the previous camera, not against current-frame travel.

This is sufficient for camera motion and stochastic Grid coverage. It is not the final answer for animated bodies, field animation, or volumetrics.

Rejection

Reject or weaken history when:

previous UV is outside the frame
travel differs too much
current or previous travel is invalid
current neighborhood color bounds do not contain the reprojected history

The first pass uses a 3x3 neighborhood min/max clamp and travel-difference weighting. Later passes add field/material ids, normals, disocclusion masks, and reactive masks.

Resolve

The resolver outputs both:

tonemapped final backbuffer color
linear HDR history for the next frame

Current blend target:

resolved = lerp(current, clamped_history, history_weight)

History weight starts conservative. Stochastic Grid coverage needs accumulation, but ghosting is worse than shimmer while the renderer lacks full velocity, material id, and reactive masks. The resolver should earn trust before taking more history.

Non-Negotiables

Do not use Playdead-style TAA as the final model; it ghosts too easily for the scene we are building.
Do not vendor Unreal TSR code. Study behavior and build a clean-room resolver.
Do not hide missing renderer signals with larger blend weights.
Do not treat volumetrics as ordinary opaque surface history.
Do not make the Grid a final-image HUD layer; it is diegetic scene UI and must respect nearer solids.

Implementation Gates

Gate 1: Owned Temporal Pass

HDR scene target
history ping-pong
camera jitter
current/previous camera reprojection
travel rejection
3x3 color clamp
final tonemap after resolve

Gate 2: Renderer Signals

material/field id buffer
normal buffer
explicit velocity for animated SDF bodies
coverage/reactive masks for stochastic Grid and future particles

Current Gate 2A implementation:

scene metadata target stores stable field id plus surface normal
history metadata ping-pongs alongside color/travel history
field id mismatch rejects history
normal mismatch weakens/rejects history
Grid stochastic surfaces write Grid travel/normal even when the current dither sample does not draw color, so the temporal resolver accumulates the diegetic surface rather than reprojecting from the solid behind it

Current Gate 2B implementation:

Self, Grid, and each orbiting planet have distinct field ids
orbiting planet hits map their current world hit back to the planet's previous center before camera reprojection
this gives analytic SDF bodies object motion without a full velocity buffer yet

Still missing:

separate field ids once the SDF field registry exists
reactive/coverage masks
richer disocclusion classification
a general velocity buffer for non-rigid fields, deformation, and future volumetrics

Gate 3: Volumetric Temporal Contract

Current Gate 3A implementation:

scene pass writes a temporal-control target separate from color/travel and field/normal metadata
temporal control currently stores:

x = reactive strength
y = stochastic/composition coverage
z = medium opacity placeholder
w = reserved

Grid coverage feeds the coverage channel
low-coverage Grid samples raise reactive strength and reduce history weight
the resolve combines reactive and coverage weights with travel, field, normal, and neighborhood-color validation

Current Gate 3B implementation:

temporal-control history ping-pongs alongside color/travel and field/normal history
resolve samples previous-frame temporal control at the reprojected UV
coverage discontinuity reduces history
medium-opacity discontinuity reduces history, even though real volumetrics do not write the channel yet
current temporal control is written into the history set for the next frame

Current Gate 3C implementation:

temporal-control history w stores accepted history age in frames
validation grows age when reprojected history survives the travel, field, normal, color, coverage, and medium-opacity checks
history authority ramps with age, so newly accepted history can contribute but does not immediately get the same weight as stable history
age resets to zero on validation failure

Current Gate 3D implementation:

travel validation uses the reprojected previous-world position's distance from the previous camera
raw current-frame travel is no longer compared directly to previous history travel, because camera motion changes the ray distance even when the world hit is stable

Current Gate 3E implementation:

stochastic Grid history bypasses opaque-surface neighborhood color clamping and current-frame color-delta rejection
Grid validation still uses field id, travel, normal, coverage continuity, reactive weight, and history age
this prevents a current dither miss from declaring the true neighborhood to be background and crushing the very history that should accumulate stochastic coverage

Current Gate 3F implementation:

stochastic Grid history uses a high-retention blend policy distinct from opaque surfaces
Grid history weight no longer scales down with current coverage or reactive strength; low coverage is exactly where temporal accumulation is needed
opaque surfaces keep the stricter color/reactive validation path
debug mode 2 should now show a smoother history signal than raw scene mode 1 when the Grid is stable

Current Gate 3G implementation:

identity/control buffers are point-loaded, not linearly sampled
field id equality is exact after point-loading, so Grid/Self/planet boundaries hard-reject mismatched history
Grid history color is sampled with the same nearest pixel used for metadata so bright foreground silhouettes do not bilinearly bleed into background Grid history
Grid max history weight is reduced from the too-soft 0.975 to 0.94

Current Gate 3H implementation:

Grid temporal coverage is now line/contour/field-line support, not broad overlay alpha
broad weather/field tint can still render through stochastic coverage, but it does not keep history age alive across the whole Grid surface
Grid history is gated by current support, so stale stochastic hits cannot billow away from the analytic line support
Grid max history weight is reduced again to 0.90 with a lower fresh-history scale for crisper line response

Current Gate 3I implementation:

Grid resolve reconstructs the analytic Grid overlay at the current world hit and uses premultiplied overlay color as the current color estimate
the scene pass may still use stochastic coverage, but the temporal resolver no longer treats the current binary hit/miss sample as the true radiance
debug mode 1 can remain noisy because it shows the raw stochastic scene; mode 0 should now use the analytic Grid estimate plus validated history

Current Gate 3J implementation:

debug mode 1 is restored to raw current scene color, before Grid analytic reconstruction
Grid final color no longer blends previous color history; it uses the analytic current Grid estimate and writes that estimate into history
history/age/weight debug modes remain useful diagnostics, but stale Grid color history no longer drives the final visible Grid

Current Gate 3L implementation:

projection jitter is scaled to zero for the live frame
visible jitter on Self/planet surfaces means the resolve is not yet hiding the sample offset well enough, so jitter stays disabled until opaque and Grid paths both prove stable
the abandoned unjittered Grid retrace was removed; with jitter disabled, Grid analytic color reconstructs from the existing current hit instead of solving the surface twice

Current Gate 3M implementation:

final presentation no longer blends temporal color history for any surface
the history, metadata, and control ping-pongs remain for debug views and for the future volumetric temporal contract
mode 0 is current-frame color after analytic Grid reconstruction and tonemapping; modes 2 through 4 show what the diagnostic temporal path would have sampled, aged, and weighted

This is intentionally still a surface/stochastic coverage contract. It prepares the resolver for volumetrics without lying that a single hit depth describes a whole participating medium.