Investors

Creative becomes predictable when you instrument it.

ImagePredict is creative experimentation infrastructure for commerce. We generate, test, and measure ad images using real products, then store the outcome-linked data. The dataset compounds into a prediction engine: rank creative before spend.

The wedge

Start as the fastest path to high-performing product and ad imagery: studio capture + retouching + controlled AI variants. Use live ad tests to generate truth.

The compounding asset

A normalized dataset linking product + image variables + audience + outcomes. Over time, we predict performance before spend. Tools are copyable. Outcome-linked data is not.

The flywheel

Every campaign increases learning velocity and reduces waste.

Generate

Real products + controlled creative variables.

Test

Live ad experiments with real spend and real audiences.

Measure

Outcomes tied back to the exact image configuration.

Normalize

Structured tables across products, images, models, campaigns, outcomes.

Predict

Rank creative before spend. Scale only high-confidence variants.

Compound

Learning becomes an appreciating asset.

Why now: Meta’s Andromeda era

Andromeda shifts Meta ads toward creative-first delivery: the platform’s automation handles most targeting and allocates impressions based on creative performance signals. Advertisers win by shipping diverse creatives quickly, measuring lift in-market, and compounding what works into a repeatable system.

What changed on Meta

  • Creative becomes the primary signal — weak or low-signal creative can be filtered out before it meaningfully reaches the auction.
  • Automation replaces manual targeting — Advantage+ style delivery shifts advertiser leverage from audience knobs to creative inputs.
  • Creative diversity is rewarded — Meta encourages multiple variations (“creative packs”) vs. relying on a single hero ad.

Why ImagePredict is advantaged

  • High-volume, on-brand variant generation from real product capture — tailored by placement, channel, and demographic context.
  • Controlled experiments in real campaigns — tie outcomes back to exact image variables to produce decision-grade learning.
  • Outcome-linked dataset → prediction — rank creative before spend and scale only high-confidence variants as the model improves over time.

Creative pack strategy

Ship meaningful concept variations (not tiny tweaks) so the algorithm has rich signals to match to users.

Faster refresh cadence

Ad fatigue can arrive in weeks; ImagePredict keeps a pipeline of validated creatives ready to deploy.

Less wasted spend

Identify losers earlier, scale winners sooner, and compound learning into repeatable playbooks.

Source: “ImagePredict: Empowering Advertisers in Meta’s Andromeda Era” (internal memo based on Meta and industry references).

Dataset schema (high level)

Stored without consumer PII. Model descriptors are categorical. Outcomes are aggregated performance metrics.

Products
  • product_id
  • category
  • material
  • price_band
Images
  • image_id
  • lens_mm
  • lighting_id
  • crop_ratio
Models
  • age_range
  • gender
  • body_type
  • skin_tone
Campaigns
  • channel
  • placement
  • audience
  • geo
Outcomes
  • impressions
  • ctr
  • cvr
  • roas_proxy

Defensibility comes from repeatable experimentation + consistent instrumentation + compounding outcome-linked data.

Why others cannot just copy it

Studios optimize for throughput and file delivery. Agencies optimize campaigns, not creative fundamentals. Generative tools lack consistent provenance and outcome linkage. ImagePredict is built around closed-loop learning.

The end state

Creative before spend: a prediction layer that ranks images for a product, channel, and audience before any budget is committed. Over time, creative becomes a measurable, forecastable input to growth.

Team

Built by operators who have shipped high-volume e-commerce creative at scale and can execute studio-to-data pipelines end-to-end. Team and bios available in the deck.

Request deck Back to homepage