LensBench · Sequential lens design workbench
Lens design is changing hands: you bring the ideas and the judgment, AI and LensBench do the hands-on work and the math. Every edit shows you exactly which numbers change and how the metrics move first, and every number we report traces back to the run that produced it.
Works with the files and AI clients you already use
It used to be that when the requirements changed, an engineer redid everything: tweak parameters, run analyses, take screenshots, copy numbers, assemble the report. Now a person does three things: state the need, confirm, sign off. AI orchestrates the rest, LensBench does the math, and every number in the report is checked against its evidence id.
What humans doAI assistantSoftware computes
Make the last surface thickness a paraxial focus solve and tell me how much the total track changes.
Dry run first, nothing written yet:
-surf 6 R=-18.39533 T=42.20778 +surf 6 R=-18.39533 T=solve(marginal_height,0) -var R1..R6 T6 +var R1..R6
Focal length and F-number are unchanged. Total track gets longer by about 0.23 mm. Apply it?
The conversation is illustrative; the diff and metrics are real output LensBench computed on the Cooke triplet reference design.
Tell the AI what you want. It dry-runs first: which numbers change and how the first-order quantities move are laid out in front of you, and nothing is written until you confirm.
Want to try a direction? Swap a glass, thicken an element, let the AI run a round of optimization. See side by side what got better and what got worse. Don't like it? Step back in one move.
Prescription diffSurf 1 T 3.26 → 4.00 · Surf 3 glass F2 → SF2 · Surf 6 T 42.44 → 45.49
| Image quality (polychromatic) | 第 2 条改动之前 | 现在 | |
|---|---|---|---|
| RMS 0°（µm） | 16.02 | 7.53 | Better |
| RMS 14°（µm） | 21.17 | 30.14 | Worse |
| RMS 20°（µm） | 23.42 | 24.88 | Worse |
| MTF 0° T @ 45/mm | 0.204 | 0.519 | Better |
| MTF 14° T @ 45/mm | 0.646 | 0.021 | Worse |
| MTF 20° T @ 45/mm | 0.07 | 0.27 | Better |
| Max distortion（%） | 0.047 | 0.166 | Worse |
Demo: three changes from three sources (AI, manual edit, CLI) land in one change log. The log, the comparison and the rollback are real LensBench output on a Cooke triplet.
When an AI reports a number, how do you know it is not made up, or out of date? Every analysis result carries an evidence id that traces back to the exact computation.
$ lb spot cooke.lens Spot diagram (µm, reference chief, 6 rings) Field RMS@0.4861 RMS@0.5876 RMS@0.6563 RMS poly GEO poly ---- ---------- ---------- ---------- -------- -------- 0 3.830 4.695 6.455 5.112 9.591 14 15.939 17.634 19.443 17.730 48.313 20 13.008 12.142 13.921 13.044 37.506 Evidence ev-7cb1d17888
The last line is the evidence id. It leads back to the model, engine and sampling parameters behind this result.
101 / 101
analytic golden references passed
Expected values come from independent formulas, never from LensBench itself
27 / 29
public designs match an independent implementation, ray by ray
Paraxial values and ray landing points reconciled after a round-trip rebuild, tolerance 1e-9. The rest are convention differences we have traced to their cause.
33 / 33
checks passed against commercial software
3 designs, with expected values computed directly by the commercial software. More settings are still being checked, and we do not claim "fully equivalent".
Figures from the validation report of Oct 21, 2026.
The desktop app is the main workbench and MCP plugs in your own AI; the web version will reopen once it moves onto the new kernel. The desktop app, the command line and MCP share one command table and one Rust kernel, so they produce identical numbers.
A 3D canvas, a table and the text stay in sync. The change timeline, review and sign-off live here. Computation runs on your machine, and so does your design.
DownloadNothing to install: try it and demo it in the browser. The web version is moving onto the same kernel as the desktop app and will reopen once that is done.
Paused while it moves to the new kernel
Claude Code, Claude Desktop, Cursor, Cherry Studio, Trae and other clients call the same command table over MCP. One command, lb mcp-config, sets it up.
Connect your AIParaxial and real ray tracing, ray aiming, every surface type and glass dispersion are computed in our own Rust kernel. The optical core is pure Rust with no third-party dependencies. Large ray batches run in parallel chunks, so the interface stays responsive while it computes in the background. Every figure and every evidence id records the engine build.
4–23×
faster local optimization
Cooke triplet 2.7 s → 0.35 s; bit-identical results when ray aiming is off
5–8×
faster across 35 analyses
Bit-identical results on all 6 reference designs
Speedups measured on a development machine on Oct 16, against a Python-based reference implementation.
From prescription to analysis, optimization and report: one text file, one set of commands.
Table, text and commands stay in live sync. Change one radius in the table and exactly one number changes in the text; comments and formatting are left alone.
Glasses match exactly as "vendor:name" and are never silently swapped for another. Use your own AGF catalogs too; all 13 dispersion formulas are supported.
Spot diagrams, ray fans, MTF, field curvature and distortion, Seidel... Every plot has a numeric page built from the same data.
Merit functions use the operand names you already know, and acceptance specs are written separately. After any change, verify has the final word.
Let an AI write the script: open a lens, tweak operands, pull analysis data, make plots, with results landing in local files. You can write it by hand too, in a numpy-style API.
A design is a text file on your computer, and the optical computation runs there too.
.lens is plain text: put it in git for diffs and rollbacks, send it to a colleague and it just opens. Export to .zmx whenever you like, so you are never locked into one tool.
.lens guideUse only the desktop app and the command line, and your design never leaves your computer. When you connect an AI, which model sees your design is decided by the client you choose. The AI can only read and write the project folder you point it to, and any text inside a design file is treated as data, never as instructions.
AI safety boundaries“The more capable the AI, the more this confirmation matters.”
The LensBench team“Read the diff, then change the design”8 classic designs ship with the app. Open any of them in one click from the desktop home page, as a starting point or a benchmark.
Double Gauss
examples/double_gauss.lens
Cooke triplet
examples/cooke_triplet.lens
Petzval
examples/petzval.lens
Telephoto
examples/telephoto.lens
Wide-angle
examples/wide_angle.lens
Fold mirror
examples/fold_mirror.lens
Parabolic mirror
examples/parabola.lens
Singlet
examples/singlet.lens
Q&A, bug reports, suggestions and numerical cross-checks. The forum bot reproduces the problems you report and tracks the fix.
Visit the forumA quickstart and user guide. The command, syntax, diagnostic-code and operand references are generated from the code, so they always match your version.
Read the docsDesign thinking and workflows: why we built it this way, and what AI-native work looks like.
Read the blog