Uni-AdGen: Unified Autoregressive Model for Personalized Image‑Text Ad Generation (CVPR 2026)

Background and Motivation

In the highly competitive e‑commerce era, high‑quality ad creatives—combined images and copy—are essential for click‑through rate improvement. Traditional pipelines suffer from three core challenges: (1) high labor cost and slow response due to reliance on designers, (2) semantic “splits” between image and text generation when using separate large language and diffusion models, and (3) noisy personalization signals from user behavior, both at the sample and modality levels.

Overall Solution: Uni‑AdGen

Uni‑AdGen proposes a unified autoregressive architecture that jointly generates visual and textual components of an ad. The system consists of two progressive stages: a General Ad Generation base model and an Personalized Ad Generation stage that adds a coarse‑to‑fine preference understanding module.

Task Definition

The goal is to produce a pair (I^{pred}, T^{pred}) that maximizes similarity to the ground‑truth pair (I^{GT}, T^{GT}) for a given user and target product. Inputs include a multi‑source history sequence of user click‑through image‑text pairs and the target product information (product background image, description, and a pool of selling points).

Architecture Overview

The model follows a standard autoregressive decoder. At each decoding step, a special control token signals whether the model should generate text ( <text> … </text>) or image ( <image> … </image>). Text tokens are decoded by a language head, while image tokens are fed to a VQ‑GAN decoder that maps discrete codes to pixels.

Training Objective

Training follows next‑token prediction for a concatenated instruction sequence s = {s_1,…,s_N}. The joint loss is L = L_{text} + L_{image} where L_{text} maximizes the probability of the next text token conditioned on previous tokens, and L_{image} does the same for image tokens.

Foreground‑Aware Control

Product background images are patchified, projected, and encoded by a DINOv2‑based foreground encoder. The resulting visual embedding is aligned to the autoregressive latent space via a simple MLP and injected additively into every fourth decoder layer, enabling the model to respect product‑specific visual constraints.

Coarse‑to‑Fine Preference Modeling

To mitigate sample‑level noise, the system first performs semantic similarity‑based sampling: from millions of historical behaviors it selects the top N candidates using product‑text similarity scores s_i with a smoothing term ε = 1e‑6. After coarse selection, a multimodal preference extractor encodes the selected image‑text pairs with two independent Transformers (visual and textual). Attention‑based relevance scores generate token‑level masks, which are refined by a differentiable Gumbel‑Softmax and Top‑K selection to suppress modality‑level noise. The fused tokens are then fed to the decoder, and style placeholders <text_ph> and <image_ph> guide final generation.

Dataset: PAd1M

A new large‑scale personalized ad dataset, PAd1M , was constructed from real e‑commerce logs. It contains 1.145 million active users, 18.92 million click events, and an average of >16 multimodal historical records per user, covering over 40 product categories. The dataset includes product foreground masks extracted by Grounded‑SAM, seller‑provided descriptions, and selling‑point pools.

Evaluation Metrics

Text quality: character‑level BLEU and ROUGE against real clicked copy.

Image quality: ImageReward, PickScore, and human preference.

Novel background metric: Product Background Similarity (PBS), trained via MoCo‑v3 on 680 k same‑product‑different‑background pairs; PBS achieved a score of 0.421, whereas CLIP and DINO were below 0.05.

Experimental Results

General Ad Generation : Uni‑AdGen achieved the best ImageReward score, ranked second on PickScore and human evaluation, and matched state‑of‑the‑art baselines on BLEU/ROUGE, demonstrating effective multimodal training.

Personalized Ad Generation : On both image and text metrics, Uni‑AdGen outperformed baselines such as Flux‑Kontext, Pigeon, Qwen‑3, and DeepSeek‑R1. Qualitative examples show that competing models either ignore user preference or suffer from sample noise, while Uni‑AdGen produces ads closely aligned with actual clicks. Ablation studies confirmed that (a) historical data, (b) product‑similarity sampling (PSS), and (c) multimodal preference extraction each contribute significantly to performance gains.

Conclusion

Uni‑AdGen demonstrates that a single autoregressive model, equipped with foreground‑aware control and a coarse‑to‑fine preference module, can generate high‑quality, personalized ad creatives at scale, and that the newly released PAd1M dataset and PBS metric provide robust benchmarks for future research.