Skip to content

armgit / 67.js

Public Imported from GitHub
0

3D engine that AI agents drive

Show the six-scene tier numbers and the Cook against Cycles comparison per sample and at equal error (Part of #890)

Nguyen Xuan An7d90bdd160b5
diff --git a/README.md b/README.md
index 378946dd..575cf7db 100644
--- a/README.md
+++ b/README.md
@@ -11,6 +11,7 @@
67js is a 3D engine for AI agents and for pages. It has a real-time rasterizer, a path tracer, addons for terrain, vegetation, characters, liquid, vehicles and media export, and an MCP server.
A scene is plain data: each part has a geometry, a matrix and a material.
+
An agent builds and changes a scene through the MCP server, looks at PNG previews and renders stills or MP4 video.
The rasterizer runs on WebGL2.
The path tracer runs on the CPU, on WebGL2 and, with the optional `packages/render-native`, on Vulkan.
@@ -38,36 +39,39 @@ The tier names are working names.
The `quality` setting that picks a tier is planned ([issue 904](https://github.com/codepawl/67.js/issues/904)), and so is refine-at-rest ([issue 903](https://github.com/codepawl/67.js/issues/903)).
Realtime-high does not exist yet ([issue 901](https://github.com/codepawl/67.js/issues/901) and [issue 902](https://github.com/codepawl/67.js/issues/902)).
-These numbers are provisional.
-They come from one scene (`lit-interior`, 960x540), medians with ranges over 3 interleaved runs, on a card shared with other jobs.
+Six scenes at 960x540, medians over 3 interleaved runs on an idle card.
FLIP is the error against a Cycles 4096-sample reference, and lower is better.
-The other five reference scenes were not run.
-The generator and raw results are in the [bench/tiers folder](https://github.com/codepawl/67.js/tree/main/bench/tiers), which lands with [pull request 931](https://github.com/codepawl/67.js/pull/931).
+Every value is the range of the per-scene medians across the six scenes.
+The full per-scene tables are in the [bench/tiers folder](https://github.com/codepawl/67.js/blob/main/bench/tiers/results/matrix-960.table.md). They land with [pull request 931](https://github.com/codepawl/67.js/pull/931), after [pull request 921](https://github.com/codepawl/67.js/pull/921) and before this one.
<!-- tiers-table:start -->
-| Tier | For | Runs on | Trade-off | 67js time | Blender equivalent and its time | FLIP error | Card memory (shared) |
-| --- | --- | --- | --- | --- | --- | --- | --- |
-| draft | The cheapest preview | Glaze, 0.5 render scale, 1 sample | Low resolution and cheap lighting | 1.20 ms per frame (0.88 to 2.43) | EEVEE 5.3 at 1 sample: 99 ms per render call (76 to 157) | Glaze 0.188, EEVEE 0.562 | Glaze +62 MiB, EEVEE +2,218 MiB |
-| realtime | Interactive frames | Glaze, default settings | Rasterized lighting | 1.11 ms per frame (1.05 to 1.24) | EEVEE 5.3 defaults: 1,127 ms per render call (716 to 1,891) | Glaze 0.197, EEVEE 0.562 | Glaze +119 MiB, EEVEE +2,503 MiB |
-| realtime-high | Better real-time frames | Planned: Glaze on WebGPU with a GI cache, temporal accumulation and upscaling ([issue 901](https://github.com/codepawl/67.js/issues/901), [issue 902](https://github.com/codepawl/67.js/issues/902)) | More VRAM | not measurable yet | none | none | none |
-| final | Finished stills | Cook native (Vulkan ray queries) | A few seconds to a few minutes per frame | 6.05 ms per sample (4.7 to 6.4), 1.9 s at 256 samples | Cycles 5.3 OptiX: 36.2 ms per sample (28.7 to 220), 12.6 s at 256 samples | At 256 samples: Cook 0.080, Cycles 0.039 | Cook +622 MiB, Cycles +733 MiB |
-| final | Finished stills | Cook webgpu (WGSL compute, software BVH) | Slower per sample than native | 34.4 ms per sample (22.6 to 60.6) | same Cycles | same as native at the same samples | +456 MiB |
-| final | Finished stills | Cook gpu (WebGL2) | Adds 4.8 to 7.9 s of shader compile | 74.4 ms per sample (57.3 to 79.2) | same Cycles | same as native at the same samples | +496 MiB |
-| final | Finished stills | Cook cpu (JavaScript, 8 threads) | No card needed, slowest | 786 ms per sample (561 to 1,022) | same Cycles | same as native at the same samples | not a card job |
+| Tier | For | Runs on | Trade-off | 67js measured | Blender measured | FLIP vs Cycles |
+| --- | --- | --- | --- | --- | --- | --- |
+| draft | A quick look while editing | Glaze, 0.5 render scale, 1 sample, light shading, 8 bit target | Aliased edges, no environment lookup, a quarter of the pixels enlarged | 0.54 to 1.04 ms a frame with the pixel read; 0.26 to 0.33 ms queued back to back | EEVEE 5.3, 1 sample: 56 to 59 ms a render call | Glaze 0.045 to 0.188, EEVEE 0.041 to 0.562 |
+| realtime | Interactive use | Glaze, default settings | Indirect light is approximate | 0.77 to 0.98 ms a frame with the pixel read; 0.35 to 0.50 ms queued back to back | EEVEE 5.3, 64 samples: 288 to 363 ms a render call | Glaze 0.030 to 0.197, EEVEE 0.042 to 0.562 |
+| realtime-high | Better interactive frames | Planned: Glaze on WebGPU with a GI cache and upscaling ([issue 901](https://github.com/codepawl/67.js/issues/901), [issue 902](https://github.com/codepawl/67.js/issues/902)) | More VRAM | not measurable yet | none | none |
+| final | Finished stills | Cook native, 256 samples (Vulkan ray queries, optional `packages/render-native`) | More samples than Cycles for the same noise | 1.73 to 4.24 ms a sample; 0.71 to 1.32 s for 256 samples with setup | Cycles 5.3 OptiX: 6.81 to 26.41 ms a sample; 3.43 to 8.46 s for 256 samples with setup | Cook native 0.019 to 0.080, Cycles 0.010 to 0.039 |
+| final | Finished stills | Cook webgpu, 256 samples (WGSL compute, software BVH, library only) | Slower per sample than native, same image | 5.88 to 17.33 ms a sample; 1.94 to 4.86 s for 256 samples with setup | same Cycles | Same as native: FLIP between the two images 0.0000 to 0.0003 |
<!-- tiers-table:end -->
-Do not read the final row as Cook being faster than Cycles.
-Cook native draws a sample about 5.4 times faster than Cycles, but it needs about 6.8 times more samples for the same error.
-At equal error the two take about the same time (0.91 to 0.95 of the Cycles time), and Cook renders about 1% brighter.
-That comparison covers one scene (`lit-interior`) at 960x540 with 8 bounces and no denoiser, on a shared RTX 3060 against Cycles 5.3 with OptiX.
-[docs/cycles-parity.md](https://github.com/codepawl/67.js/blob/main/docs/cycles-parity.md) has the method and its limits. It lands with [pull request 921](https://github.com/codepawl/67.js/pull/921).
-
-The Glaze times are milliseconds per frame.
+Glaze runs the realtime tier in under about 1 ms a frame at 960x540 on these scenes (0.77 to 0.98 ms), measured with a pixel read.
The EEVEE times are one full background render call, including scene sync, with EEVEE ray tracing off and no global illumination.
-The two columns are not a frame rate comparison.
-The draft tier is not faster than realtime here. At 960x540 a Glaze frame costs about as much as the browser pixel read, and the ranges overlap.
+They are not a frame rate comparison.
+The draft tier is not faster than realtime here. At 960x540 a Glaze frame costs about as much as the browser pixel read, so the queued figures are the ones that differ.
+
+Cook native draws a sample 3.8 to 6.2 times faster than Cycles 5.3 OptiX, but needs more samples for the same noise.
+On the three noise-limited scenes (lit-interior, spot-area and light-linking) it needs 6.6 to 19.0 times Cycles' samples for the same error.
+That makes it 1.05 to 4.02 times Cycles' time for the sampling alone, and 0.87 to 2.52 times with setup included.
+
+On metal, glass and outdoor-hdri the error falls slower than 1/N between 256 and 1024 samples, which points to bias.
+Cook does not reach Cycles' error there, so no equal-error ratio is given.
+Cook webgpu draws a sample 3.3 to 4.1 times slower than native.
+
+Cook gpu (WebGL2) and Cook cpu are not in this run.
+Cook cpu needs no card, so VRAM does not apply.
+[docs/cycles-parity.md](https://github.com/codepawl/67.js/blob/main/docs/cycles-parity.md) has the Cook against Cycles method on one scene.
The experimental real-time path tracer is not the realtime-high tier.
On the WebGPU backend it draws a frame in 5.6 ms.
@@ -77,9 +81,8 @@ It ends at 0.056 after 256 frames.
The WebGL2 fallback takes 6.7 ms per frame and 6.1 s for the first frame, which includes shader compile.
It runs at 2 bounces against an 8-bounce reference.
-Machine: RTX 3060 12 GB, Intel i7-14700KF, headless Chrome 157 on D3D11, Node 24, Blender 5.3, measured 2026-11-11.
-The card was shared with other jobs in every run, so the ranges are wide.
-Card memory is the card-wide change while the job ran, an upper bound.
+Machine: RTX 3060 12 GB (driver 621.14), Intel i7-14700KF, Chrome 157 headless on D3D11, Node 24, Blender 5.3, measured 2026-11-11.
+The card was idle: each run waited for at most 15 percent utilisation, and 0 of 162 runs were contended.
## Showcase
diff --git a/changes/890-renderer-copy.md b/changes/890-renderer-copy.md
index 606e506d..ea93682a 100644
--- a/changes/890-renderer-copy.md
+++ b/changes/890-renderer-copy.md
@@ -1,2 +1,2 @@
-- The README, home page, `llms.txt`, package keywords and plugin text name each renderer backend (Glaze on WebGL2, Cook on CPU, WebGL2 and Vulkan, with opt-in WebGPU) (#890)
-- The README and home page show a provisional measured tier table for one scene, with realtime-high marked as planned (#890)
+- The README, home page, `llms.txt`, package keywords and plugin text name each renderer backend (Glaze on WebGL2, Cook on CPU and WebGL2, and on Vulkan with the optional `packages/render-native`, with opt-in WebGPU) (#890)
+- The README and home page show the measured tier table across six scenes on an idle card, with realtime-high marked as planned and Cook against Cycles stated per sample and at equal error (#890)
diff --git a/index.html b/index.html
index 0d0260b3..47363f26 100644
--- a/index.html
+++ b/index.html
@@ -4,11 +4,11 @@
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<meta name="theme-color" content="#09070f" />
- <meta name="description" content="67js is a 3D engine that AI agents drive: the Glaze rasterizer on WebGL2, the Cook path tracer on CPU, WebGL2 and Vulkan, GLB, PNG and MP4 export, and an MCP server. Install with npm i @codepawlhq/67js." />
+ <meta name="description" content="67js is a 3D engine that AI agents drive: the Glaze rasterizer on WebGL2, the Cook path tracer on CPU and WebGL2 (Vulkan with the optional render-native package), GLB, PNG and MP4 export, and an MCP server. Install with npm i @codepawlhq/67js." />
<meta property="og:type" content="website" />
<meta property="og:site_name" content="67js" />
<meta property="og:title" content="67js · 3D engine that AI agents drive" />
- <meta property="og:description" content="A WebGL2 rasterizer, a path tracer on CPU, WebGL2 and Vulkan, and an MCP server in one npm package." />
+ <meta property="og:description" content="A WebGL2 rasterizer, a path tracer on CPU and WebGL2 (Vulkan with an optional package), and an MCP server in one npm package." />
<meta property="og:url" content="https://67js.codepawl.com/" />
<meta property="og:image" content="https://67js.codepawl.com/hero-lab/poster-b-port.webp" />
<meta property="og:image:width" content="1440" />
@@ -85,7 +85,7 @@
</div>
<div class="copy">
<h1>3D engine that<br />AI agents drive</h1>
- <p class="lede">Glaze draws real-time frames on WebGL2. Cook path traces on the CPU, WebGL2 or Vulkan. 24 MCP tools drive them.</p>
+ <p class="lede">Glaze draws real-time frames on WebGL2. Cook path traces on the CPU or WebGL2, and on Vulkan with the optional render-native package. 24 MCP tools drive them.</p>
<div class="hero-actions">
<a href="/docs.html#getting-started" data-icon="book">Get started</a>
<a href="https://www.npmjs.com/package/@codepawlhq/67js"><span class="package-logo package-logo-npm brand-npm" aria-hidden="true"></span>npm</a>
@@ -169,28 +169,26 @@ renderer.render({ eye: [58, 31, 84], target: [-8, 3, -18] });</code></pre>
</div>
<div class="agent-cards">
<a href="/mcp.md" data-icon="cube"><h3>24 MCP tools</h3><p>Create, edit, explode, blend and render over stdio.</p></a>
- <a href="/path-tracer.md" data-icon="code"><h3>Glaze + Cook</h3><p>Real-time frames on WebGL2, or path-traced stills on CPU, WebGL2 or Vulkan.</p></a>
+ <a href="/path-tracer.md" data-icon="code"><h3>Glaze + Cook</h3><p>Real-time frames on WebGL2, or path-traced stills on CPU or WebGL2, and on Vulkan with the optional render-native package.</p></a>
<a href="/agent-review.md" data-icon="image"><h3>Visual review</h3><p>Compare labeled views and normals from one snapshot.</p></a>
</div>
</section>
<section id="tiers" class="tiers" aria-labelledby="tiers-title">
<h2 id="tiers-title">Quality tiers, measured</h2>
- <p>Provisional numbers from one scene (lit-interior, 960x540) on an RTX 3060 shared with other jobs. Medians of 3 runs, so the ranges are wide. FLIP is the error against a Cycles 4096-sample reference, lower is better. Tier names are working names, and the <code>quality</code> setting that picks one is planned.</p>
+ <p>Six scenes at 960x540 on an idle RTX 3060. Each value is the range of the per-scene medians across the scenes. FLIP is the error against a Cycles 4096-sample reference, lower is better. Tier names are working names, and the <code>quality</code> setting that picks one is planned.</p>
<div id="tiers-table">
<ul class="tier-list">
- <li class="tier-row tier-head" aria-hidden="true"><span>Tier</span><span>For</span><span>Runs on</span><span>Trade-off</span><span>67js</span><span>Blender</span><span>FLIP error</span></li>
- <li class="tier-row"><span data-label="Tier">draft</span><span data-label="For">The cheapest preview</span><span data-label="Runs on">Glaze, 0.5 scale, 1 sample</span><span data-label="Trade-off">Low resolution, cheap lighting</span><span data-label="67js">1.20 ms per frame</span><span data-label="Blender">EEVEE 1 sample: 99 ms per render call</span><span data-label="FLIP error">Glaze 0.188, EEVEE 0.562</span></li>
- <li class="tier-row"><span data-label="Tier">realtime</span><span data-label="For">Interactive frames</span><span data-label="Runs on">Glaze, defaults</span><span data-label="Trade-off">Rasterized lighting</span><span data-label="67js">1.11 ms per frame</span><span data-label="Blender">EEVEE defaults: 1,127 ms per render call</span><span data-label="FLIP error">Glaze 0.197, EEVEE 0.562</span></li>
- <li class="tier-row"><span data-label="Tier">realtime-high</span><span data-label="For">Better real-time frames</span><span data-label="Runs on">Planned: Glaze on WebGPU with a GI cache and upscaling (issues 901 and 902)</span><span data-label="Trade-off">More VRAM</span><span data-label="67js">not measurable yet</span><span data-label="Blender">none</span><span data-label="FLIP error">none</span></li>
- <li class="tier-row"><span data-label="Tier">final</span><span data-label="For">Finished stills</span><span data-label="Runs on">Cook native, Vulkan</span><span data-label="Trade-off">A few seconds to a few minutes per frame</span><span data-label="67js">6.05 ms per sample</span><span data-label="Blender">Cycles OptiX: 36.2 ms per sample</span><span data-label="FLIP error">At 256 samples: Cook 0.080, Cycles 0.039</span></li>
- <li class="tier-row"><span data-label="Tier">final</span><span data-label="For">Finished stills</span><span data-label="Runs on">Cook webgpu, opt-in</span><span data-label="Trade-off">Slower per sample than native</span><span data-label="67js">34.4 ms per sample</span><span data-label="Blender">same Cycles</span><span data-label="FLIP error">same as native at the same samples</span></li>
- <li class="tier-row"><span data-label="Tier">final</span><span data-label="For">Finished stills</span><span data-label="Runs on">Cook gpu, WebGL2</span><span data-label="Trade-off">Adds 4.8 to 7.9 s of shader compile</span><span data-label="67js">74.4 ms per sample</span><span data-label="Blender">same Cycles</span><span data-label="FLIP error">same as native at the same samples</span></li>
- <li class="tier-row"><span data-label="Tier">final</span><span data-label="For">Finished stills</span><span data-label="Runs on">Cook cpu, 8 threads</span><span data-label="Trade-off">No card needed, slowest</span><span data-label="67js">786 ms per sample</span><span data-label="Blender">same Cycles</span><span data-label="FLIP error">same as native at the same samples</span></li>
+ <li class="tier-row tier-head" aria-hidden="true"><span>Tier</span><span>For</span><span>Runs on</span><span>67js</span><span>Blender</span><span>FLIP error</span></li>
+ <li class="tier-row"><span data-label="Tier">draft</span><span data-label="For">A quick look while editing</span><span data-label="Runs on">Glaze, 0.5 scale, 1 sample</span><span data-label="67js">0.54 to 1.04 ms a frame</span><span data-label="Blender">EEVEE 1 sample: 56 to 59 ms a render call</span><span data-label="FLIP error">Glaze 0.045 to 0.188, EEVEE 0.041 to 0.562</span></li>
+ <li class="tier-row"><span data-label="Tier">realtime</span><span data-label="For">Interactive use</span><span data-label="Runs on">Glaze, defaults</span><span data-label="67js">0.77 to 0.98 ms a frame</span><span data-label="Blender">EEVEE 64 samples: 288 to 363 ms a render call</span><span data-label="FLIP error">Glaze 0.030 to 0.197, EEVEE 0.042 to 0.562</span></li>
+ <li class="tier-row"><span data-label="Tier">realtime-high</span><span data-label="For">Better interactive frames</span><span data-label="Runs on">Planned: Glaze on WebGPU with a GI cache and upscaling (issues 901 and 902)</span><span data-label="67js">not measurable yet</span><span data-label="Blender">none</span><span data-label="FLIP error">none</span></li>
+ <li class="tier-row"><span data-label="Tier">final</span><span data-label="For">Finished stills</span><span data-label="Runs on">Cook native at 256 samples, Vulkan with the optional render-native package</span><span data-label="67js">1.73 to 4.24 ms a sample</span><span data-label="Blender">Cycles OptiX: 6.81 to 26.41 ms a sample</span><span data-label="FLIP error">Cook native 0.019 to 0.080, Cycles 0.010 to 0.039</span></li>
+ <li class="tier-row"><span data-label="Tier">final</span><span data-label="For">Finished stills</span><span data-label="Runs on">Cook webgpu at 256 samples, library only</span><span data-label="67js">5.88 to 17.33 ms a sample</span><span data-label="Blender">same Cycles</span><span data-label="FLIP error">Same as native</span></li>
</ul>
</div>
- <p>At equal error, Cook native and Cycles take about the same time. Cook native draws a sample about 5.4 times faster than Cycles but needs about 6.8 times more samples. The Cook time is 0.91 to 0.95 of the Cycles time, and Cook renders about 1% brighter. This is one scene at 960x540 with 8 bounces and no denoiser, against Cycles 5.3 with OptiX. <a href="https://github.com/codepawl/67.js/blob/main/docs/cycles-parity.md">docs/cycles-parity.md</a> has the method.</p>
- <p>Glaze times are per frame. EEVEE times are one full background render call, including scene sync, with EEVEE ray tracing off and no global illumination. They are not a frame rate comparison. Raw results: <a href="https://github.com/codepawl/67.js/pull/931">the tier benchmark, pull request 931</a>.</p>
+ <p>Glaze runs the realtime tier in under about 1 ms a frame (0.77 to 0.98 ms), measured with a pixel read. Cook native draws a sample 3.8 to 6.2 times faster than Cycles 5.3 OptiX, but needs more samples for the same noise: 6.6 to 19.0 times Cycles' samples on the three noise-limited scenes (lit-interior, spot-area, light-linking), which is 1.05 to 4.02 times Cycles' time for the sampling alone. On metal, glass and outdoor-hdri Cook does not reach Cycles' error, because the error there is limited by bias, so no ratio is given. Cook cpu needs no card, so VRAM does not apply.</p>
+ <p>Glaze times are per frame. EEVEE times are one full background render call, including scene sync, with EEVEE ray tracing off and no global illumination. They are not a frame rate comparison. Per-scene tables: <a href="https://github.com/codepawl/67.js/blob/main/bench/tiers/results/matrix-960.table.md">bench/tiers on main</a>.</p>
</section>
<section id="documentation" class="documentation-links" aria-labelledby="documentation-title">
diff --git a/public/llms.txt b/public/llms.txt
index fbec5025..50b5c8b7 100644
--- a/public/llms.txt
+++ b/public/llms.txt
@@ -1,6 +1,6 @@
# 67js
-> A 3D engine that AI agents drive. Glaze draws real-time scenes on WebGL2, and Cook path traces stills on the CPU, WebGL2 or Vulkan. PNG, MP4 and GLB output comes from the same scene.
+> A 3D engine that AI agents drive. Glaze draws real-time scenes on WebGL2, and Cook path traces stills on the CPU or WebGL2, and on Vulkan with the optional render-native package. PNG, MP4 and GLB output comes from the same scene.
67js has an MCP server (`67js-mcp`, 24 tools) that builds, edits, looks at and renders scenes from JSON recipes.
The same engine is a JavaScript library for pages and Node. The library API is in the guide below.
diff --git a/scripts/generate-llms.mjs b/scripts/generate-llms.mjs
index 15d67119..9135ab8e 100644
--- a/scripts/generate-llms.mjs
+++ b/scripts/generate-llms.mjs
@@ -61,7 +61,7 @@ export function buildLlmsText(tools) {
const toolLines = tools.map(tool => `- \`${tool.name}\`: ${summarize(tool.description ?? '')}`).join('\n');
return `# 67js
-> A 3D engine that AI agents drive. Glaze draws real-time scenes on WebGL2, and Cook path traces stills on the CPU, WebGL2 or Vulkan. PNG, MP4 and GLB output comes from the same scene.
+> A 3D engine that AI agents drive. Glaze draws real-time scenes on WebGL2, and Cook path traces stills on the CPU or WebGL2, and on Vulkan with the optional render-native package. PNG, MP4 and GLB output comes from the same scene.
67js has an MCP server (\`67js-mcp\`, ${tools.length} tools) that builds, edits, looks at and renders scenes from JSON recipes.
The same engine is a JavaScript library for pages and Node. The library API is in the guide below.
diff --git a/skills/render-67js-media/SKILL.md b/skills/render-67js-media/SKILL.md
index ee6add93..c82ab8a6 100644
--- a/skills/render-67js-media/SKILL.md
+++ b/skills/render-67js-media/SKILL.md
@@ -1,6 +1,6 @@
---
name: render-67js-media
-description: Export a 67js scene to PNG stills (Glaze, the WebGL2 rasterizer, or Cook, the path tracer on CPU, WebGL2 or Vulkan), render passes or an MP4 through the MCP render job tools (scene_render_submit, scene_render_status, scene_render_result) of the npm package @codepawlhq/67js. Use when an agent must render an image or video of a scene, play a camera or light animation to video, cut shots with a timeline, add a score, write depth or object-ID passes, path trace a still, choose GPU or CPU and an encoder, or stay inside memory limits. Also for "render ảnh 3D", "render video 3D", "xuất mp4", "export 67js", "render bị lỗi bộ nhớ", "chất lượng MSAA". Scene building is the build-67js-scene skill.
+description: Export a 67js scene to PNG stills (Glaze, the WebGL2 rasterizer, or Cook, the path tracer on CPU or WebGL2, and on Vulkan with the optional render-native package), render passes or an MP4 through the MCP render job tools (scene_render_submit, scene_render_status, scene_render_result) of the npm package @codepawlhq/67js. Use when an agent must render an image or video of a scene, play a camera or light animation to video, cut shots with a timeline, add a score, write depth or object-ID passes, path trace a still, choose GPU or CPU and an encoder, or stay inside memory limits. Also for "render ảnh 3D", "render video 3D", "xuất mp4", "export 67js", "render bị lỗi bộ nhớ", "chất lượng MSAA". Scene building is the build-67js-scene skill.
license: MIT
---
diff --git a/src/site.css b/src/site.css
index bec8cf20..05aabc64 100644
--- a/src/site.css
+++ b/src/site.css
@@ -382,7 +382,7 @@ body.site > main > section { scroll-margin-top: 96px; }
.site .tiers p { color: var(--muted); max-width: 70ch; }
.site .tiers p a { text-decoration: underline; }
.site .tier-list { display: grid; gap: 6px; margin: 0; padding: 0; list-style: none; }
-.site .tier-row { display: grid; grid-template-columns: 0.9fr 1.1fr 1.5fr 1.4fr 1.1fr 1.6fr 1.5fr; gap: 12px; padding: 12px 16px; border-radius: var(--panel-radius); background: var(--bg); font-size: 13.5px; }
+.site .tier-row { display: grid; grid-template-columns: 0.9fr 1.1fr 1.8fr 1.2fr 1.6fr 1.6fr; gap: 12px; padding: 12px 16px; border-radius: var(--panel-radius); background: var(--bg); font-size: 13.5px; }
.site .tier-row > span { min-width: 0; overflow-wrap: anywhere; }
.site .tier-row > span:first-child { color: var(--text); font-weight: 600; }
.site .tier-head { background: none; padding-top: 0; padding-bottom: 0; color: var(--muted); font-size: 12.5px; }