Canopy Labs
of Bangalore
Canopy Labs is an applied research lab making taste measurable, so the models and agents that create produce work that feels right, not only work that is correct.
Read the manifestoGeneration is free now. Judgment is the scarce thing, and we are building it.
Generation is free now. Anyone can make a thousand images, a hundred ads, ten films before lunch.
So the scarce thing is no longer making. It’s knowing what to throw away.
The shift
The challenge is knowing what to make. And how to make it great.
One prompt and boom: a dashboard, a short story, a landing page, a game, a slide. Your timeline is full of it. And it all looks the same.
The generic, soulless sameness. That is slop.
The problem
01 / Pretraining
Next-token prediction learns a probability-weighted blend of everything. For 2+2 the average is optimal. For design, the average is mediocre, and distinctive work is by definition low-probability.
02 / Preference tuning
Reward models average annotators of different tastes and learn what nobody dislikes, not what anyone loves. RLHF-tuned models produce measurably less diverse outputs than their base models.
03 / Rewards
Math and code got good on verifiable rewards. There is no function that returns 1 for good design and 0 for bad. So subjective domains never got to ride that flywheel.
Kirk et al. (2024); Padmakumar & He (2024): RLHF reduces output diversity against base models. Read the research note
What we believe
Good in Mumbai isn’t good in New York. Good for one brand is wrong for another. There is no single good, only good for this: this brief, this brand, this audience, this moment.
A model trained on the average of everywhere is from nowhere.
Our insight
Taste is the bar of quality in the absence of correctness
Seeing many things.
Noticing patterns.
Deciding what you like.
Deciding what you don’t.
If it’s pattern recognition, it can be decomposed. If it can be decomposed, it can be measured. If it can be measured, it can be improved.
Down-selection
Every great creative director, editor and art director is defined less by what they make than by what they reject. Three directions killed, one approved, with a reason. That reason is the most valuable thing in creative work, and today it disappears the moment the file is saved. We’re here to keep it.
Models learn from the vote. Quality lives in the reasons.
What we build
Raise the floor and the ceiling, at training and at inference.
For labs giving their models taste, vision and design capability. We evaluate where a model breaks, then craft the exact data shape, quantity and distribution to fix it.
Talk to us about a programInfrastructure for app-layer companies, coding agents and creative tools: context, evaluation and verification, so they output better, more beautiful, more personal and on-brand work.
See the OnBrand APIHow it works
We turn something like design into the shape of a problem like code: measurable, decomposable, visualisable. Every judgment routes to the cheapest signal that can carry it.
Every T3 judgment becomes tomorrow’s T2 model.
Programmatic, deterministic. Cheap at scale.
↓ what code can’t judge
Trained on tastemaker data. Judges against a reference.
↓ what models can’t judge yet
The tails: where experts genuinely disagree.
The network
100+
Tastemakers in the network: designers, art directors, copywriters
Vetted tastemakers nominate others. Trust propagates through socially validated judgment. They design, grade AI outputs and review each other’s work. Small groups of expert judges, made dramatically more efficient by our tooling.
Live · used daily by agencies
Two agents that do the work. Every pick, reject and reason teaches a brand brain that stays the brand’s.
# Roots — inside the canvas > "3 Diwali ad cuts, on-brand, 9:16" roots.plan → brain.recall · image.gen ×3 · video.gen ‖ assist → approve spend? ✓ # brain.recall → living brand memory (excerpt) { "brand": "your-brand", "voice": "warm, regional, never salesy", "palette": [ "#1b1b1b", "#f2c14e", "#fff8ec" ], "never": [ "stock smiles", "purple gradients" ], "last_reject": "too glossy, not us" } # Soil — on the timeline soil.trim(clip_2, "00:03.2") soil.caption(lang: "hi-IN") · soil.export("9:16")
Illustrative flow. Assist mode approves before spend.
OnBrand can help your agent with things like brand extraction
Explore the OnBrand APISo far
20+
Design pilots
6
Products live, in beta or alpha
5
People. Capital-efficient by design
The cure for slop isn’t better generation. It’s better judgment. Here is the argument, in full.
Team
A small team of researchers, makers and engineers, working in person in Bangalore. We’ve spent our careers studying how people choose, and we know the people whose choices are right.
