Primary target surface
These rendering routes are generated from canonical source. Native-WebGPU acceptance remains separate from entrypoint availability.
Bloom scene-linear HDR in Three.js WebGPU/TSL. Use when choosing optical full-scene glare, selective or hybrid contributors, transparent contribution blending, or viewport, exposure, and performance gates for BloomNode.
These rendering routes are generated from canonical source. Native-WebGPU acceptance remains separate from entrypoint availability.
This skill contributes to the following cross-skill owner graphs.
Source hashes, claim verdicts, promoted same-lab media, fixed routes, exact tier contracts, and current limitations.
Bloom is a camera response to bright HDR signal: energy above a threshold scatters into a wide kernel. The signal chain is the contract — bloom samples scene-linear HDR before tone mapping:
$$L_{bloom} = K_{blur} * \max(L_{scene} - T, 0), \qquad L_{final} = \operatorname{tonemap}(L_{scene} + s\,L_{bloom})$$The blur is a mip pyramid: progressive downsample + upsample accumulation approximates a large Gaussian at a fraction of the cost, with radius scaling per mip level. Selective bloom routes emissive contribution through an MRT channel so only authored emitters glow:
$$L_{bloom} = K * (E_{mrt}), \qquad E = \text{emissiveNode output, scene-relative units}$$The skill's core rule: an object's brightness hierarchy is authored in scene units (sun ≫ plasma ≫ LED); bloom reveals that hierarchy, it must never be the object.
Only schema-v2 labs with accepted runtime and evidence contracts appear here. Other source directories remain visible through the demo registry without being promoted to runnable proof.
Every image identifies what it proves. Page screenshots demonstrate the published presentation only; generated inputs demonstrate asset channels only; rendering acceptance still requires same-lab readback and a schema-v2 bundle.
0 published imagesThe complete SKILL.md as loaded by agents, rendered verbatim.
Treat bloom as the broad tail of an imaging point-spread function. Its closest built-in optical input is all scene-linear HDR radiance. Selective contribution is an art-direction branch, not an automatic quality upgrade.
Capture scene-linear HDR, false-color pre-tone luminance, and the final image with bloom disabled. Repair the source when highlights were clamped or when silhouette, material, or lighting identity disappears without bloom.
This step is complete when the bloom-off scene remains readable and the input is unclamped scene-linear HDR in a declared working basis.
| Contract | Signal | Admission test |
|---|---|---|
| optical glare from emission, direct response, reflection, transmission, sky, and sun | full scene HDR | a luminance threshold and knee meet contributor error limits |
| named surfaces glow differently from equally bright radiance | selective emissive contribution | full-scene threshold cannot meet those limits and the MRT delta fits |
| optical highlights plus deliberate boost | scene HDR plus selective boost | both components are required and the shared attachment is charged once |
| source scene fails without glare or all bloom tiers miss a gate | defer bloom | source is repaired or budget/viewport changes |
Stabilize exposure before comparing branches. Inspect both false-positive and false-negative bloom energy, including bright mirrors and transmissive surfaces.
This step is complete when one source branch is named, its omissions are accepted, and its contribution error and target-device cost meet declared limits.
Use one HDR scene traversal for both full-scene and selective bloom. Full-scene
bloom reads the scene output. Selective bloom adds an emissive MRT output to
the same pass and gives it BlendMode(MaterialBlending) so transparent
contributions follow the material blend state.
In r185, material-level mrtNode merging can lose the operative blend-mode map.
Keep the canonical selective path on the regular emissiveNode; use a separate
measured contribution pass or source-verified custom merge only when visible
emission and bloom contribution must diverge.
Compose vec4(scene.rgb + bloom.rgb, scene.a). Verify transparent overlap in
both insertion orders and under the chosen premultiplied/straight-alpha policy.
This step is complete when scene traversal count is known, every contribution uses the intended depth/sort/blend/alpha state, overlapping contributors accumulate correctly, and scene alpha is unchanged by bloom.
When implementing either graph or transparent selective contribution, read r185 graphs and transparent blending.
Name the threshold domain:
Convert threshold and soft-knee width together when exposure or calibration
changes. Treat radius as cross-mip spread, not a physical radius. Validate
minimum and maximum viewport/DPR. Stock r185's five-level chain requires:
floor(bloomScale * min(drawingBufferWidth, drawingBufferHeight)) >= 16
This step is complete when threshold units and update owner are explicit, the halo passes endpoint views, and the deepest mip remains valid at every shipping extent.
When tuning spread, scale, or a nonstandard optical footprint, read BloomNode PSF and work.
Use this default order:
stable scene-linear HDR
-> temporal resolve, when present
-> excluded transparent/refractive layers
-> pre-bloom meter tap
-> bloom extraction, blur, and RGB add
-> adapted exposure
-> tone map, grade, and one output conversion
Selective input that bypasses temporal resolve still needs a stability decision
for subpixel or discontinuous emission. Keep UI and diagnostics outside HDR
bloom unless the visual contract includes them. After changing the active
output graph, set RenderPipeline.needsUpdate = true.
This step is complete when bloom has one owner, every admitted meter, exposure, tone map, alpha operation, and output conversion has one owner, and the bloom-off graph is genuinely reachable.
When coupling threshold to exposure or selecting final output ownership, read Exposure, output, and lifecycle.
Use $threejs-exposure-color-grading for exposure-coupled thresholds, tone
mapping, grading, and output conversion. Use $threejs-image-pipeline when the
scene pass, MRT signals, transparent ordering, or final graph is shared with
other effects.
Measure warmed paired graphs: scene alone, full-scene bloom, scene with contribution MRT, and selective/hybrid bloom where applicable. Report physical pixels, scale, attachment formats, target inventory, transparent screen coverage, sustained timing, and marginal time. Reduce bloom scale first for a pixel-bound miss, then remeasure; full-scene input can remove a costly full-resolution contribution attachment. Disable bloom when the minimum-mip, memory, thermal, or marginal-time gate still fails.
Replace the output node for the disabled path, mark the pipeline dirty, and
dispose BloomNode and any exclusive contribution resources after final GPU
use. Recreate evidence after renderer/device loss or format generation change.
This step is complete when every shipping tier passes fixed captures and sustained target-device budgets, bloom-off timing loses the bloom work, and repeated enable/disable, resize, and disposal cycles stabilize resources.
When calculating exact internal storage/work or diagnosing a failed gate, read Budget and acceptance.
Preserved concept proxies and generated-asset previews. They are excluded from primary completion counts and link to the canonical lab through the schema-v2 registry.