{"id":"7a5609e8-ea7d-4f98-92b2-a2d9d970d05b","arxiv_id":"2501.14316","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A four-stage product-centric framework automatically creates advertising posters from a product photo, taglines, and target size, with better visual-quality scores and about 3 percent higher click-through rates in Taobao A/B tests.","lead":"Researchers from Alibaba built a four-stage system that turns a product photo, marketing taglines, and a target size into a finished advertising poster, by generating a matching background and arranging text and product placement automatically. The system was tested in live Taobao ad slots and improved click-through rates by about 3 percent, suggesting that automated ad design can produce commercially useful results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported Taobao CTR gains (3.02% and 3.03%) are given without confidence intervals, p-values, or impression/click counts, so the only unfiltered real-world evidence for the central claim may be statistical noise.","rationale":"The reader's CONDITIONAL verdict is appropriate. I considered the dataset-filtering concern directly: Section 3 excludes hard cases, and Section 6 admits failures on product angle/lighting, multiple foregrounds, and more than five taglines. However, an online A/B on 5,000 real products would largely settle the generalization question if its statistics were reported properly. The paper does not provide the numbers needed to make that judgment, so the load-bearing weak point is not the filtering per se but the unquantified reliability of the only unfiltered evaluation. If the A/B confidence interval excludes zero, the central claim is supported despite the filtered offline data; if it includes zero, the strongest real-world evidence fails. No other internal inconsistency was found: the four-stage architecture, training scheme, ablations, and user study are coherent and directionally consistent. The user study (7 users, 200 samples) is small but supports the visual-appeal claim. Thus my recommendation is unchanged: CONDITIONAL, with the added requirement to report A/B uncertainty explicitly.","tokens_in":14315,"tokens_out":8243,"duration_ms":79465,"concrete_test":"Ask the authors for the per-product impression and click counts from both Taobao A/B scenarios, or reconstruct them from the served traffic logs. Compute a cluster-robust 95% confidence interval for the CTR lift, bootstrapping by product to account for intra-product correlation. If the interval excludes 0 in both scenarios, the online generalization claim survives; if it includes 0, the central claim loses its only unfiltered deployment evidence and the paper should be asked to present corrected or additional online evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the 3.02% and 3.03% CTR increases in Section 5.5 reflect a real effect rather than sampling variation. The A/B description gives no number of impressions or clicks, no confidence intervals, no significance tests, and no adjustment for intra-product correlation; it only says 5,000 products were selected, 5% of main traffic was used, and data were gathered for one month. With typical e-commerce CTRs, a ~3% relative lift can easily fall inside the noise band for such a design, especially if the effective number of experimental products is small. This matters because the offline evidence is built on the filtered PITA/PIL distributions (Section 3 excludes messy backgrounds, plain colors without shadows, close-ups, poorly presented taglines, and cluttered elements). The A/B is therefore the only evidence that the gains transfer to unfiltered advertiser inputs. If the A/B confidence interval includes zero, the paper's strongest external-validation claim is unsupported; the user study and FID results remain, but they are confined to a filtered test distribution and to author-constructed baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents T-Stars-Poster, a product-centric framework for automatic advertising image generation. Given a product foreground image, taglines, and a target canvas size, the framework runs four stages: prompt generation with a fine-tuned VLM, joint layout prediction for graphic (tagline, underlay, logo) and nongraphic (background objects) elements with a second VLM, background image generation with an SDXL-based inpainting model that receives prompt, layout, and foreground controls, and a rule-based graphics rendering step. The authors collect and label two datasets, PITA (38,017 images) and PIL (12,247 images), and report offline comparisons against two self-built pipelines, against prompt- and layout-generation baselines, and against layout-controlled inpainting models, plus a user study with 7 users on 200 samples and online A/B tests in two Taobao scenarios. The headline result is a lower FID (37.524 vs. 43.221 for the stronger baseline), favorable layout metrics, user win rates of 74-94%, and CTR increases of 3.02% and 3.03% in the two online scenarios.","tokens_in":14507,"tokens_out":7591,"duration_ms":65189,"significance":"T-Stars-Poster is an industrially motivated and internally coherent system. Its main strengths are the joint prediction of graphic and nongraphic layouts, the use of Deep Layer Control and LoRA adaptation to stabilize the layout-and-inpainting training, the creation and public release of the PITA dataset, and the deployment of the system in live Taobao traffic. If the online gains are real, the paper demonstrates a practical solution to a hard problem. However, the evidence base has three important gaps: the A/B results are reported without confidence intervals, sample sizes, or significance tests (Section 5.5); the training and test distributions are filtered to exclude hard advertising cases (Section 3), and the authors' own Section 6 limitations confirm that angle/lighting, multiple foregrounds, and many-tagline cases remain unsolved; and the full-pipeline comparisons use only author-constructed baselines, with P&R re-implemented by the authors (Section 5.3). These gaps do not invalidate the system, but they materially narrow the scope of the central claim as currently worded.","major_comments":[{"comment":"The online A/B test is the only evidence that the method generalizes beyond the filtered PITA/PIL distributions, and it is reported without any uncertainty quantification. The text states that 5,000 products were selected, 5% of main traffic was used, data were collected for one month, and CTR increased by 3.02% and 3.03% in two scenarios, but it gives no impression counts, click counts, confidence intervals, or p-values, and it does not account for intra-product correlation. Under typical e-commerce CTRs, a roughly 3% relative lift with 5,000 products can easily be within sampling noise, especially if the effective number of independent experimental units is smaller. Please report per-scenario sample sizes, CTRs with confidence intervals, and a significance test (or a permutation test over products), and clarify whether the 5,000 products pass the Section 3 filtering criteria or are a random sample of unfiltered advertiser inputs.","section":"Section 5.5"},{"comment":"The PITA and PIL datasets explicitly exclude images with messy backgrounds, plain colors without shadows, unattractive stickers, close-ups, poorly presented taglines, and cluttered elements. Because every offline result in Tables 1-5 is computed on this filtered distribution, the reported FID, layout metrics, and user preferences are conditional on a curated subset of advertising images. The authors should quantify how representative this subset is of the deployed scenarios: for example, report the fraction of real advertiser submissions that satisfy the filtering criteria, or hold out an unfiltered set and report per-category performance. Without this, the abstract's claim of automatically generating advertising images from basic inputs overstates the demonstrated scope, and the limitation statement in Section 6 suggests that several common cases (lighting/angle mismatch, multiple foregrounds, more than five taglines) are not handled.","section":"Section 3"},{"comment":"The headline full-pipeline comparison is against two author-constructed baselines, Pipeline 1 and Pipeline 2, and the foreground-based baseline P&R is re-implemented by the authors because the original code is not public. While the authors are transparent about this, the risk of unintended bias is real: the baselines' prompts are generated by GPT-4o with no fine-tuning, the layout method PosterLlama operates on a fixed image, and the same rendering module is applied to all. Please provide a more detailed specification of the P&R re-implementation, state which version or checkpoint of PosterLlama is used, and make the comparison pipeline code available so the baselines can be checked. At a minimum, report the number of human-evaluation samples and the inter-rater agreement for the aesthetic rankings used in Table 1.","section":"Section 5.3, Table 1"},{"comment":"The user study is very small: 7 users, 200 randomly selected test samples, pairwise comparisons. The reported win rates of 74-94% have no confidence intervals, no per-user breakdown, and no measure of inter-rater agreement. Given that this study is one of the two pieces of evidence for the \"more visually appealing\" claim, please report the full preference matrix, a binomial or mixed-effects confidence interval for each win rate, and the number of comparisons per pair. This is a fixable presentation issue, but without it the reader cannot tell whether the preference differences are statistically reliable.","section":"Section 5.4, Table 5"}],"minor_comments":[{"comment":"SceneBooth [4] is cited as the first method to combine inpainting and layout control in training, but it is not included in the comparisons of Table 4. If SceneBooth is not publicly available or is not applicable to advertising images, please say so explicitly; otherwise add it as a baseline.","section":"Section 2.2 and Table 4"},{"comment":"The RKBR representation is described in prose as [x, y, h*r1/r2, h] but the example JSON in Figure 3 shows a raw arithmetic expression in the bbox field; please make the notation consistent and define r1 and r2 unambiguously with respect to the foreground and canvas aspect ratios.","section":"Section 4.3"},{"comment":"The ED Score is defined as the FID between CLIP features of prompts and ground-truth prompts; please specify the CLIP model variant, the feature extraction layer, and the number of prompt samples used for the FID computation.","section":"Section 5.2"},{"comment":"The paper states the PITA dataset is available at GitHub but does not state the same for PIL. Please clarify whether PIL will also be released, and specify the exact train/test splits for both datasets so that future comparisons use identical splits.","section":"Section 3"},{"comment":"In the example tagline translation, \"Made in good materialsand not break within 5 yeas\" contains a typo (\"yeas\" should be \"years\", and there is a missing space before \"and\"); please correct the figure or its caption.","section":"Figure 3"},{"comment":"The limitations paragraph is candid and useful. Please consider adding quantitative evidence for the stated failure modes (for example, the percentage of test cases with lighting mismatch or more than five taglines) so that readers can gauge the operational scope of the method.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"For the editor: this is an industrial systems paper with a useful dataset release. The main risk is not the architecture, which is plausible and internally consistent, but the statistical backing of the live experiment and the generality of the filtered datasets. I would recommend asking for a careful revision of Section 5.5 with the raw A/B numbers, and possibly a statistical review. Also, the paper's relationship to SceneBooth [4] should be clarified, since the authors cite it as the closest prior work but do not compare against it experimentally."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here is my honest take on T-Stars-Poster. It is a well-engineered industrial framework for product-centric advertising image design, and the central claim—that jointly predicting graphic and nongraphic layouts before background generation produces better posters—is plausible and mostly supported. The genuinely new pieces are JPGNL, the ratio-keeping bbox representation (RKBR), and the class-conditioned occlusion strategy (CCLP). The two datasets (PITA, PIL) are a real contribution, and the ablations for DLC and LAT are carefully done. The paper is honest about its limitations: it cannot adjust product angle/lighting, multiple foregrounds are treated as one, and layout degrades with more than five taglines.\n\nThe soft spots are real but not fatal. First, the online A/B in Section 5.5 reports only percentage CTR lifts (3.02% and 3.03%) with no sample sizes, confidence intervals, or significance tests. That is the only unfiltered evidence, and the stress-test concern is fair: a 3% lift over 5,000 products with 5% traffic may or may not be noise. The authors should report impression/click counts and at least a standard error. Second, the full-pipeline baselines are author-constructed. Re-implementing P&R is reasonable, but the comparison remains self-built. Third, the dataset filtering (excluding messy backgrounds, plain colors, close-ups, etc.) likely removes the hard cases; the framework may work well on the curated distribution and less well on the long tail. These caveats weaken but do not overturn the central claim.\n\nWho should read this? Researchers and engineers working on e-commerce image generation, layout generation, and design automation. The paper is not a scientific breakthrough, but it is a competent industrial contribution with new data and a clear method. It deserves a serious referee, not a desk reject. I would ask the authors to strengthen the A/B reporting and to discuss the filtering bias more explicitly before publication.","headline":"A well-engineered industrial framework for product-centric ad design with genuinely new joint layout prediction; the A/B evidence is weaker than reported.","tokens_in":15052,"tokens_out":2390,"would_cite":true,"duration_ms":21732,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"T-Stars-Poster generates advertising posters from a product image, taglines, and a target size, and reports better visual appeal and higher click-through rates.","keywords":["advertising image generation","layout generation","text-to-image","inpainting","visual language model","e-commerce","A/B testing","product-centric design"],"falsifier":"Run T-Stars-Poster on a random sample of unfiltered product photos from a large e-commerce catalog—including messy backgrounds, close-ups, and poorly composed shots—and measure CTR or user preference against a fixed baseline. If the quality gap and the 3% CTR gain disappear on this unfiltered set, the claimed advantage would be shown to depend on the curated training distribution rather than the framework's design.","tokens_in":14107,"feed_emoji":"🖼️","tokens_out":12072,"duration_ms":83359,"temperature":0.7,"pith_summary":"This paper claims that advertising images can be generated automatically from three basic inputs—a product foreground image, marketing taglines, and a target size—by decomposing the task into four sequential stages: prompt generation, layout generation, background image generation, and graphics rendering. The central premise is that jointly planning the placement of the product, the graphic elements (taglines, underlays, logos), and the nongraphic elements (objects described in the background prompt) before generating the background resolves the conflicts that arise when these components are designed independently. To support this, the paper introduces two datasets with over 50,000 labeled advertising images and trains separate expert models for prompt and layout generation, plus an SDXL-based inpainting model that accepts layout, prompt, and foreground controls simultaneously. The reported results show that the framework outperforms a stronger baseline pipeline in FID (37.524 vs. 43.221), wins user preference in 74% to 94% of paired comparisons, and increases click-through rate by 3.02% and 3.03% in two online A/B tests. If correct, this means that high-quality advertising creatives can be produced at scale without human designers.","feed_headline":"Automated pipeline turns product shots into ad posters and lifts CTR 3%","feed_subtitle":"It takes one product photo and taglines and outputs a full poster, beating baselines in 74–94% of user comparisons and lifting CTR by 3%.","key_machinery":"The load-bearing mechanism is the combination of joint layout prediction (JPGNL) with deep-layer control and LoRA-adapted training in the image generator. JPGNL uses a ratio-keeping bbox representation (RKBR) to preserve the product's aspect ratio and a class-conditioned layout prediction (CCLP) to allow or forbid tagline occlusion depending on product type, so that one model can allocate space for the product, text, decorative underlays, and background elements in a single pass. The generator, built on SDXL, injects layout embeddings through gated self-attention layers placed only in the deep UNet blocks (deep layer control, DLC) and is trained with a LoRA adaptation strategy (LAT) to bridge the new layout path with the pretrained model; this allows prompt, layout, and foreground inpainting to be combined without destabilizing training. What this machinery does is coordinate the three competing signals—foreground identity, textual message, and background scene—so that none is sacrificed for the others.","core_discovery":"The central discovery is that making the layout the joint product of foreground, text, and background constraints, rather than an afterthought, is what enables high-quality automatic advertising posters. In the proposed framework, a fine-tuned visual language model first writes a background prompt that matches the product's shape and setting; a second VLM-based module then outputs a single JSON layout that assigns coordinates to the product, the tagline blocks, decorative underlays, and background objects; and an SDXL-based inpainting generator uses that layout as control alongside the prompt and the product foreground. The claim is that this sequence—especially the joint layout prediction and the layout-conditioned generation—produces images that are more harmonious, more readable, and more attractive to users than existing pipelines that generate the background first and place text later.","pith_inferences":["The framework's success on filtered datasets suggests that its real-world applicability depends on expanding the training distribution to include messy backgrounds, close-ups, and other hard cases; a natural next step is to test on unfiltered product feeds.","The joint layout idea is not specific to advertising; it could be applied to other design tasks such as infographics, slides, or UI design where text and content must coexist.","The paper's own limitations—single foreground, up to five taglines, and fixed lighting—indicate that the framework is best suited to controlled product shots; extending it to arbitrary angles, lighting, or multiple foregrounds remains open.","The reported A/B test gains are measured on one platform and product mix; whether they generalize to other marketplaces or categories is an open question that could be answered by replication."],"forward_implications":["Advertising images can be produced at scale from a product photo, taglines, and a target size, which could substantially reduce the manual design effort in e-commerce.","The reported 3% CTR lift in two e-commerce scenarios implies that automatically generated posters can improve recommendation engagement, not just offline quality metrics.","Because the layout is predicted jointly for graphic and nongraphic elements, the framework can adapt product placement and tagline position to different aspect ratios and product shapes.","The two released datasets provide labeled training data for future work on product-centric advertising generation."],"supporting_citations":[{"why":"provides the SDXL architecture that the image generator is built on.","marker":"[28]"},{"why":"supplies the visual language model backbone used for both prompt and layout generation.","marker":"[9]"},{"why":"introduces gated self-attention for injecting layout embeddings into the UNet.","marker":"[19]"},{"why":"provides low-rank adaptation (LoRA) used to stabilize training when adding layout control.","marker":"[16]"},{"why":"supplies the inpainting control mechanism that integrates the product foreground into generation.","marker":"[44]"},{"why":"is the strongest background-based layout baseline the framework is compared against.","marker":"[32]"},{"why":"is the foreground-based baseline re-implemented for comparison.","marker":"[20]"}],"fun_headline_variants":["Joint layout model automates ad poster design, boosting CTR","From product shot to ad poster: new framework wins user tests","T-Stars-Poster: four-stage pipeline generates ads humans prefer","Layout-first approach yields ad posters that outperform baselines","Automated ad design: joint layout and text placement lifts CTR 3%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's training and test data deliberately exclude images with messy backgrounds, plain colors without shadows, unattractive stickers, close-ups, poorly presented taglines, or cluttered elements, so the reported quality and CTR gains may not extend to the full, unfiltered population of real advertising inputs.","fun_headline_variants_meta":{"raw":{"variants":["Joint layout model automates ad poster design, boosting CTR","From product shot to ad poster: new framework wins user tests","T-Stars-Poster: four-stage pipeline generates ads humans prefer","Layout-first approach yields ad posters that outperform baselines","Automated ad design: joint layout and text placement lifts CTR 3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000708,"raw_usage":{"total_tokens":3185,"prompt_tokens":936,"completion_tokens":2249,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2162}},"tokens_in":552,"tokens_out":2249,"duration_ms":13723,"temperature":1.0,"reasoning_tokens":2162,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:14:55.779768+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run T-Stars-Poster on a random sample of unfiltered product photos from a large e-commerce catalog—including messy backgrounds, close-ups, and poorly composed shots—and measure CTR or user preference against a fixed baseline. If the quality gap and the 3% CTR gain disappear on this unfiltered set, the claimed advantage would be shown to depend on the curated training distribution rather than the framework's design.","supporting_citations":[],"review_version":1}