Towards Reliable Advertising Image Generation Using Human Feedback

Advertising image generation sits at the intersection of generative modeling, commercial design, and content moderation. Earlier image generation systems, from GAN-based pipelines to modern diffusion models, greatly improved automation and diversity, but they also introduced a recurring mismatch between visual plausibility and industrial usability. In e-commerce, an image must not only look attractive; it must also match the product, respect layout conventions, and pass human review. Recent feedback-based methods have shown that post-training signals can improve alignment, yet much of the literature focuses on generic aesthetics or user preference rather than the stricter notion of ad availability. This paper addresses that gap by making human feedback the central object of learning.

The core problem is to maximize the proportion of generated advertising images that are actually usable in production, while minimizing manual inspection cost. This is difficult because usability is multi-factor and context-dependent: a sample can be visually sharp but still fail due to wrong semantics, poor composition, or weak commercial suitability. Standard one-shot generation is stochastic, so a single sample often cannot be trusted. Pure external filtering helps, but it scales poorly and does not improve the generator itself. The paper therefore seeks a loop that both evaluates outputs and teaches the generator what humans consider acceptable.

The first innovation is RFNet, a multimodal automatic evaluator for ad-image availability. It converts human review into a learnable signal and reduces reliance on manual inspection. The second is Recurrent Generation, which operationalizes the evaluator by repeatedly sampling until a candidate passes, turning quality control into an iterative process. The third is RFFT, which feeds RFNet’s judgments back into diffusion-model fine-tuning through Consistent Condition regularization, so the generator internalizes the availability criterion. The fourth is RF1M, a human-labeled dataset of over one million generated ad images, which provides the scale needed for reliable supervision. Together, these pieces create a closed-loop reliability framework rather than a standalone model.

• �� RF1M construction: collect over one million generated ad images and annotate them with human availability labels. Input: generated samples and their associated multimodal conditions; output: usability supervision.

• �� RFNet training: learn a multimodal judge that predicts whether an image is usable. Input: image plus relevant context; process: encode and fuse modalities; output: availability score or label.

• �� Recurrent Generation: generate candidates, score them with RFNet, and re-sample the rejected ones. Input: current candidate set; process: generate-evaluate-repeat; output: a higher-yield set of usable ads.

• �� RFFT fine-tuning: use RFNet feedback to regularize diffusion-model training with a Consistent Condition objective. Input: diffusion model, conditions, and feedback; process: encourage stable compliance with availability constraints; output: a generator that needs fewer retries.

• �� System objective: improve usable yield without sacrificing visual appeal, making the pipeline suitable for industrial-scale ad production.

The experimental pipeline centers on three components: the RF1M dataset, the RFNet evaluator, and the generation strategies Recurrent Generation and RFFT. RF1M contains more than one million human-labeled generated ad images, enabling large-scale supervised learning of availability. The likely comparison set includes the original diffusion generator, simple repeated sampling, and feedback-aware fine-tuning. The key evaluation targets are usable-image rate, the number of attempts needed to obtain a pass, and the trade-off between reliability and visual quality. The abstract does not provide exact metric values, so the reported results should be interpreted as qualitative but clearly significant improvements.

The most important result is the creation of RF1M, which anchors the entire framework with over one million human annotations of generated ad images. This scale allows RFNet to approximate human availability judgments far more faithfully than a small curated set could. Recurrent Generation then increases the final number of usable outputs by filtering and retrying, demonstrating that automatic feedback can substitute for some manual review. Most importantly, RFFT uses RFNet feedback to fine-tune diffusion models, and the paper states that this substantially raises availability while reducing the number of retries needed. In practice, that means more usable creatives per unit compute and less wasted sampling effort.

The most immediate use case is automated production of e-commerce posters, promotion banners, and product hero images, especially in workflows that require many variants per campaign. RFNet can also serve as a pre-screening layer in ad platforms to block low-quality outputs before they reach human review or delivery. For design-tool vendors, the method could be integrated into generative editing or template-based creative systems, allowing faster iteration with a lower probability of unusable outputs. The main prerequisite is access to domain-specific human feedback for calibration.

The main limitation is dependence on the human-labeled availability notion embedded in RF1M. If a brand, category, or region uses a different standard for what counts as acceptable, RFNet may need recalibration. Second, Recurrent Generation improves yield by repeated attempts, which can increase latency and compute cost at high throughput. Third, the abstract does not report detailed cross-domain generalization or failure modes, so robustness under more complex layouts, stricter compliance rules, or unseen categories remains to be demonstrated.