acceptodds
Under review as a conference paper at ICLR 2027

VPBRANDLOOP: AIGC ADVERTISING IMAGE GENERATION WITH CROSS-SCENE VANISHING-POINT CONSTRAINTS

Abstract

Computational advertising requires scalable creative generation that preserves scene structure and brand identity while supporting reliable pre-deployment quality control. We present VPBrandLoop (Vanishing-Point and Brand-conditioned Closed-Loop Advertising Generation), a framework for multi-scene AIGC advertising image generation that addresses perspective drift, under-utilized visualbrand conditions, and the absence of verifiable output-side geometric control. A precomputed dominant vanishing point (VP) and perspective attributes are injected into a latent diffusion model through multi-scale geometric residuals, while product and logo references are encoded as visual brand tokens and fused through decoupled cross-attention. Training uses brand-region-weighted diffusion reconstruction, spatial-gating alignment, and auxiliary brand-discriminative supervision. At inference time, a line-segment detector (LSD) and RANSAC re-estimate the dominant VP of each generated candidate; spatial-consistency quality gating and confidence-weighted geometric feedback then drive candidate resampling. We additionally use a CLIP-based zero-shot protocol to evaluate the multimodal semantic retrievability of brands, product categories, and visually observable attributes. On 1,000 AdsGeo-10K test images evaluated across 12 methods, VPBrandLoop achieves an N-VPE of 0.2558 and an SCPR of 16.9%. Relative to ControlNet,the strongest external baseline on the corresponding metrics, N-VPE decreases by 20.3% and SCPR increases by 4.2 percentage points; both paired differences remain significant after Holm correction. The evidence therefore supports a primary advantage in preserving dominant perspective consistency between the input scene and the generated advertisement.

Primary area:
generative models
Keywords:
vanishing-point constraints, AIGC advertising image generation, spatial consistency verification, visual brand conditioning, Generative Engine Optimization (GEO), multimodal semantic retrievability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aLhbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes VPBrandLoop to preserve scene structure and brand identity for advertising image generation. A precomputed dominant vanishing point, perspective attributes, product and logo references are input into latent diffusion models through multi-scale geometric residuals and visual tokens. Then the authors use a line-segment detector and RANSAC to get geometric feedback and do sampling.

Critical strengths
  1. The paper focuses on a critical research problem. The geometric consistency of advertising image generation has been neglected in most previous works.
  2. The paper compares multiple baselines, although some of them are outdated.
Critical weaknesses
  1. The generation quality is not good in Figure 2. It seems that the proposed VPBrandLoop cannot generate some images with correct geometric relationships. The cans look like they are directly pasted onto the backgrounds.
  2. The authors only provide a few visual examples and there is no human study. So it is unclear whether the generation quality is good.
  3. The Brand Accuracy is 100 in Table 1 but 27 in Table 2. Although these results are from different subsets, the gap is so large.
  4. Removing "input VP Injection" even improves the SCPR in Table 2.
  5. The proposed method is like a combination of existing methods.
Questions for authors
  1. Please see the Weaknesses.
  2. Could the authors provide more visualization examples?
  3. Could the authors compare the proposed method with a simple Best-of-N baseline?
Reviewer_aSGzOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

VPBrandLoop combines scene geometry, visual brand conditioning, and output-side verification for advertising image generation. Precomputed vanishing-point features enter a latent diffusion model through multi-scale residuals, while separate brand attention and region-weighted objectives encourage product fidelity. An LSD+RANSAC verifier scores generated candidates and guides selection or resampling for up to three rounds. On 1,000 AdsGeo-10K test images, the method improves normalized vanishing-point error and spatial-consistency pass rate over ControlNet, without consistently leading on semantic retrieval or VP detection rate.

Critical strengths
  1. The framework explicitly distinguishes supplying a geometric condition from verifying that a generated image satisfies it.

  2. Direct training objectives, stop-gradient diagnostics, and inference-time control are distinguished.

  3. The paper acknowledges non-significant gains, closed-set brand-recognition limitations, and the unresolved independent benefits of input VP injection and residual feedback.

Critical weaknesses
  1. Selection and evaluation share the same potentially biased verifier. Both candidate selection and final geometric scores rely on LSD+RANSAC, so better scores may reflect adaptation to the verifier rather than correct perspective. Improved geometry scores after removing brand conditioning indicate sensitivity to product edges and package text. Moreover, preserving a global scene VP does not establish correct product orientation, scale, or contact relationships. Independent geometry annotations and product-level evaluation are needed.

  2. The baseline comparison is insufficiently competitive. The evaluation omits strong image-editing systems such as Qwen-Image-Edit, Nano Banana, GPT-Image, and FLUX.2.

  3. Large differences between main and ablation results remain unexplained. Brand accuracy and semantic recall are 100.0% and 81.4% in the main test, but only 27.0% and 21.5% for the full model in the ablation batch; the reported FID also changes from 23.96 to 174.97. These batches should not be directly subtracted, but the magnitude of the differences requires explicit checkpoint, brand-set, conditioning, and scoring details.

  4. The nominal split contains 8,000 training images, but training uses 4,000. Sample selection.

Questions for authors

Please refer to the Weakness 1 2 3 and 4

Reviewer_JxxeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on generating ad images by placing brands (e.g., via their assets) in an organic/non-branded image. The setup is well motivated and the authors propose a vanishing point (VP) focused approach leveraging diffusion models. There are novel components in the approach (e.g., sampling loops based on verifications), and the offline results look positive. The main concern in the paper is that while it is motivated from an ads setup, there is no human evaluation or results from a real ads setup/platform. Without these results its difficult to validate the proposed approach.

Critical strengths
  1. Well motivated setup
  2. Diffusion based approach which focuses on vanishing points to improve quality of ad images
  3. Positive offline result using model based/automatically generated metrics.
Critical weaknesses
  1. The major concern is around lack of human evals and results from a real ads platform. Without these results we cannot gauge whether online users will engage in the generated ad images.
  2. The writing in the paper can improve, e.g., the authors can explain their focus on vanishing points in a better manner.
Questions for authors

Are there any human evals or results from a real world ads setup which indicate that online users will engage better with ads from the proposed approach compared to other baselines?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.