OptiCover: A Scalable and Multi-Surface Framework for Cold-Start Visual Optimization under Semantic Constraints

Guillaume Guy, Mihajlo Grbovic, Joe Mills, Selen Uguroglu, Liwei He, Pratiksha Kadam, Nalini Kartha, Qian Aria LiarXiv 2026 · arXiv

TL;DR

OptiCover is a scalable framework developed by Airbnb to solve the "cold-start" problem of selecting the best cover photo for millions of listings. By combining lightweight vision models, large-scale semantic grounding via Vision-Language Models (VLMs), and context-aware ranking, the system delivers a 1.5% engagement lift and a 0.35% increase in bookings, translating to millions of dollars in incremental revenue.

What Problem Does This Paper Address?

In e-commerce, the first image a user sees—the cover photo—is the most critical factor for engagement. On Airbnb, hosts manually select one static cover photo. However, this approach has three major flaws:

  1. The Cold-Start Regime: Over 90% of listing photos have never been used as a cover, meaning there is zero historical engagement data to help an algorithm decide if they are good.
  2. Context Sensitivity: A photo that works well for general browsing on the HomePage might be a poor choice for a user specifically filtering for "homes with pools."
  3. Semantic Misalignment: A visually stunning photo of a nearby sunset or a local landmark might get clicks, but if it doesn't represent the actual property, it misleads guests and degrades trust.

OptiCover aims to dynamically select the most engaging, contextually relevant, and truthful cover photo at a scale of hundreds of millions of images.

Figure 1: Airbnb Search with Pool as filter before and after applying OptiCover

Methods

The authors propose a modular architecture that balances offline efficiency with online exploration:

  • Candidate Generation: Instead of evaluating every photo for every listing, the system uses proxies (e.g., photos previously used as covers or those semantically similar to high-performing ones) to create a high-quality pool.
  • Vision Heads: To handle billion-scale scoring with low latency, they use a frozen SigLIP-2 backbone. They train shallow MLP "heads" on top of these embeddings to predict click-through rate (CTR) and identify semantic attributes (e.g., "is there a pool?").
  • Listing Grounding Validator: To ensure "truthfulness," they use a VLM (GPT-4o mini) to compare a candidate photo against the listing's text description and captions. This acts as a conservative filter to reject misleading images.
  • Exploration Logging Module (ELM): The system uses an $\epsilon$-greedy policy to show new photos to a small percentage of users. This collects the unbiased data needed to train the CTR models and overcome the cold-start problem.
Figure 2: OptiCover architecture

Datasets

The framework relies on several data sources:

  • Airbnb Internal Data: Hundreds of millions of listing photos and billions of user engagement signals (clicks, bookings).
  • VLM/Expert Labels: A dataset of 1,000 expert-annotated photo-listing pairs was used to benchmark the Grounding Validator.
  • SigLIP-2 Embeddings: Pre-trained embeddings served as the foundation for all visual tasks.

Key Insights and Ideas

  • Action-Space Shaping is King: The authors found that simply expanding the pool of eligible high-quality photos was a primary driver of performance. If your candidate set is poor, no amount of algorithmic sophistication will save you.
  • Context Trumps Raw CTR: On the Search page, a non-contextual model that just picked the "prettiest" photo increased clicks but did not increase bookings. However, a contextual model that showed a pool photo when a user filtered for "pools" significantly increased bookings. This suggests that relevance to intent is more important than general aesthetic appeal.
  • Offline Priors Beat Bandits in Cold Start: In extreme cold-start scenarios, standard bandit approaches like Thompson Sampling converged too slowly. A strong offline model trained on proxy data performed better initially, though the gap narrowed as more online data was collected.
  • Semantic Grounding is Essential for Safety: Using a VLM to ensure photos actually represent the property reduced the rate of non-representative photos by 5.3x compared to vision-only approaches.

Key Results

OptiCover was evaluated through multiple large-scale A/B tests:

  • HomePage: Achieved a 1.5% lift in CTR.
  • Search (Contextual): Delivered a 0.35% increase in bookers, which represents a significant impact on Airbnb's bottom line.
  • Vision Head Accuracy: The custom vision heads significantly outperformed zero-shot CLIP baselines, achieving a macro F1 of 0.96 for room type classification and 0.90 for amenity detection.

Limitations and Future Work

While successful, the authors note several areas for improvement:

  • Personalization: Currently, the system is contextual (based on filters/surface) but not personalized to individual guest preferences (e.g., a family might prefer different photos than a solo business traveler).
  • Joint Optimization: The system currently optimizes the cover photo in isolation. Future work could involve jointly ranking the listings themselves and their photos to maximize overall marketplace health.
  • UI Constraints: The current system can only swap the cover photo; it cannot yet reorder the entire gallery or change the layout of the listing card.

Summary updated

Abstract

Selecting an effective cover photo is a critical yet understudied problem in visual merchandising for e-commerce marketplaces. On Airbnb, hosts upload many photos depicting different rooms, amenities, and points of interest, but must manually choose a single static cover photo that is reused across heterogeneous product surfaces. This creates substantial challenges: over 90\% of listing photos have no engagement labels, leading to a severe cold-start regime; guest preferences vary across surfaces and search intents, and visually captivating photos may be semantically misaligned with the underlying listing. At Airbnb scale, this corresponds to hundreds of millions of photos and hundreds of thousands of new uploads per day. We propose \emph{OptiCover}, a scalable framework for cold-start visual optimization under semantic constraints and multi-surface engagement signals. The framework combines (i) a photo candidate generator that identifies candidate photos through various proxies, (ii) lightweight vision heads applied to frozen CLIP embeddings for large-scale semantic scoring, (iii) a cross-modal grounding validator based on a vision-language model (VLM) that enforces semantic representativeness, and (iv) an Exploration Logging Module (ELM) that orchestrates the collection of randomized engagement data across heterogeneous product surfaces and contextual search intents. We evaluate OptiCover through extensive offline studies and multiple online experiments across HomePage, and Search surfaces. OptiCover is now deployed in production and performs tens of millions of photo swaps daily. Online results show a 1.5\% engagement lift on the HomePage and a 0.35\% increase in bookers on Search, generating millions of dollars in incremental revenue. Our findings suggest that action-space shaping and semantic constraints are primary drivers of performance.
BibTeX
@misc{guy2026opticover,
  author = {Guillaume Guy and Mihajlo Grbovic and Joe Mills and Selen Uguroglu and Liwei He and Pratiksha Kadam and Nalini Kartha and Qian Aria Li},
  title = {{OptiCover: A Scalable and Multi-Surface Framework for Cold-Start Visual Optimization under Semantic Constraints}},
  year = {2026},
  eprint = {2611.12727},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.12727},
}
Figure