acceptodds
Under review as a conference paper at ICLR 2027

FidPoster: End-to-End Product Poster Generation with Fine-grained Product Detail Fidelity

Abstract

Fine-grained Product Detail Fidelity refers to preserving a reference product’s packaging text, textures, and colors in generated posters. Existing models often distort these details, compromising commercial usability despite visually appealing designs. Our analysis shows that VAE reconstruction degrades fine-grained details, highlighting the need for direct supervision of decoded appearance. During sampling, the absence of explicit fidelity guidance can further allow product details to drift from the reference. These observations motivate FidPoster, an end-to-end framework for Fidelity-preserving product Poster generation that combines pixel-space fidelity learning with reference-guided trajectory correction. During training, Pixel-Space Fidelity Activation (PFA) directly supervises decoded product regions to align optimization with visible detail preservation. During inference, Energy-Guided Velocity Correction (EVC) uses bidirectional correspondence energy to measure inconsistent and missing details relative to the reference. Gradients of this energy correct the sampling velocity to suppress detail drift without requiring externally provided product masks at inference. We also construct FidPoster14K with 14K product-poster pairs and FidPoster-Bench, a 1,000-sample benchmark comprising an 800-sample Fine-grained Fidelity setting for controlled fidelity evaluation and a 200-sample Independent Reference setting for assessing generalization to independently sourced products. Experiments across both settings demonstrate consistent improvements in product-detail fidelity while maintaining competitive promotional-text accuracy and poster quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Product Detail Fidelity, Fine-grained

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_2A3xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents FidPoster to preserve fine-grained product details in poster generation, motivated by experiments showing that VAE compression degrades high-frequency information. PFA applies pixel-space fidelity supervision and jointly trains a reference-aware localizer; EVC uses this localizer and correspondence energy to correct sampling velocities without external masks. The authors also introduce FidPoster14K and FidPoster-Bench, achieving state-of-the-art results in automatic and human evaluations.

Critical strengths
  1. The paper is well-written and the overall idea is straight-forward.
  2. FidPoster-Bench has good efforts. It evaluates both paired fidelity and unpaired generalization.
Critical weaknesses
  1. The baseline may be unfair. All baselines are zero-shot, while FidPoster is fine-tuned on in-domain data. Plain LoRA SFT (Table 4: P-Text 82.09, SSIM 0.794) already beats every baseline in Table 2. The real gain of PFA+EVC over SFT is modest (+3.4 P-Text, +0.069 SSIM). SFT should appear in Table 2.
  2. A user study is necessary to assess the perceptual quality and overall effectiveness of the generated results.
  3. Comparisons with strong proprietary models, such as GPT Image 2, could provide a more meaningful assessment of the proposed method's performance.
Questions for authors

Please refer to weakness part.

4 comments
AuthorsOpenReview ↗

Response to reviewer 2A3x

W1: Fairness of Baseline Comparisons and Contributions of PFA and EVC

We have revised Table 2 in the main paper to directly include the results of Standard SFT and PFA-only, and adjusted the original description of Improvement accordingly. The revised Table 2 clearly distinguishes between two types of comparisons: the overall performance comparison of existing methods under zero-shot settings, and the controlled comparison of Standard SFT, PFA, and EVC using the same Qwen-Image-Edit backbone and training data. Tables 4–5 further analyze the contributions of individual fidelity components.

To further distinguish the gains from PFA and Standard SFT, we investigate whether these improvements transfer to product images from independent sources. To this end, we supplement the Independent Reference evaluation set with human evaluations of Standard SFT and PFA-only. The results are shown below:

MethodProduct Fidelity ↑Layout & Composition ↑Visual Quality ↑Commercial Usability ↑Overall ↑
Baseline4.6434.1114.0093.8954.165
+ Standard SFT4.6774.2804.1804.1324.317
+ PFA (w/o EVC)4.7704.3064.2024.2124.373
FidPoster (Ours)4.8384.3264.2444.2354.411

On the Product Fidelity metric, which is the primary focus of this work, Standard SFT achieves a score of 4.677. Introducing PFA increases the score to 4.770, and further enabling EVC raises it to 4.838. This demonstrates that both PFA and EVC improve product detail preservation beyond standard SFT, even on the Independent Reference evaluation set, which contains no paired target posters.

Combined with the controlled automatic evaluation in Table 2 and the newly added human evaluation on Independent Reference, these results allow us to more clearly distinguish the contributions of each stage: standard SFT improves the model's basic product generation capability, PFA further enhances product detail preservation through decoded pixel-space supervision, and EVC provides additional reference-aware fidelity correction during inference. These results demonstrate that the product fidelity gains of the complete FidPoster method cannot be attributed solely to standard SFT.

AuthorsOpenReview ↗

Response to reviewer 2A3x

W2: Human Evaluation of Perceptual Quality and Overall Effectiveness

A systematic human evaluation has already been conducted in Table 3 of the main paper. We evaluate 200 Independent Reference product samples across four dimensions: Product Fidelity, Layout & Composition, Visual Quality, and Commercial Usability. These dimensions assess the consistency between generated products and reference images, the visual presentation of the posters, and their practical applicability. The existing human evaluation results are as follows:

MethodProduct Fidelity ↑Layout & Composition ↑Visual Quality ↑Commercial Usability ↑Overall ↑
BAGEL2.7891.5691.6291.3721.840
FLUX.24.0442.5502.4331.9812.752
IC-Custom3.7732.2592.2711.8442.537
OmniGen22.2831.3071.4371.2751.576
PosterOmni3.9203.7853.6313.3543.673
Qwen-Image-Edit-25094.6434.1114.0093.8954.165
Step1X-Edit4.4611.5541.5591.4502.256
FidPoster (Ours)4.8384.3264.2444.2354.411

To further address the reviewer’s concerns regarding Perceptual Quality and Overall Effectiveness, we additionally introduce these two independent human evaluation dimensions based on the original evaluation. Perceptual Quality primarily measures the overall perceptual quality of generated posters, including visual naturalness, detail presentation, and generation artifacts. Overall Effectiveness comprehensively considers product presentation, visual appearance, and the fulfillment of design objectives to evaluate the overall effectiveness of the generated posters.

Considering the workload associated with human evaluation, we select three representative comparison methods: Qwen-Image-Edit-2509, PosterOmni, and Step1X-Edit. Among them, Qwen-Image-Edit-2509 is the base model adopted by FidPoster, PosterOmni is a representative method specifically designed for product poster generation, and Step1X-Edit is an image editing model that performs well in automatic product detail evaluation. These three methods provide comparisons from the perspectives of the base model, specialized poster generation methods, and image editing methods, respectively.

The results of the newly added Perceptual Quality and Overall Effectiveness evaluations are presented below:

MethodPerceptual Quality ↑Overall Effectiveness ↑
PosterOmni3.8283.614
Qwen-Image-Edit-25094.5114.083
Step1X-Edit3.5652.367
FidPoster (Ours)4.7314.356

Among the three selected baselines, FidPoster achieves the highest scores in both Perceptual Quality and Overall Effectiveness. Specifically, FidPoster achieves a Perceptual Quality score of 4.731, outperforming Qwen-Image-Edit-2509 and PosterOmni by 0.220 and 0.903, respectively. Its Overall Effectiveness score reaches 4.356, exceeding the two methods by 0.273 and 0.742, respectively. These results demonstrate that FidPoster not only improves the fidelity of fine-grained product features, such as packaging text, logos, colors, and local textures, but also maintains good overall perceptual quality and complete poster generation performance.

AuthorsOpenReview ↗

Response to reviewer 2A3x

W3: Comparison with the Strong Closed-Source Model GPT-Image-2

Considering the relatively high inference cost of GPT-Image-2 and the workload associated with human evaluation, we conduct an additional comparison on the 200 Independent Reference samples in FidPoster-Bench, selecting two representative human evaluation dimensions: Product Fidelity, which measures the fine-grained product fidelity that is the primary focus of this work, and Perceptual Quality, which evaluates the overall visual perceptual quality of generated posters.

MethodProduct Fidelity ↑Perceptual Quality ↑
GPT-Image-24.7464.749
FidPoster (Ours)4.8384.731

The results show that although FidPoster is built upon the open-source Qwen-Image-Edit model, it achieves performance comparable to GPT-Image-2 on both evaluation dimensions. In terms of Perceptual Quality, FidPoster obtains a score of 4.731, only 0.018 lower than GPT-Image-2's score of 4.749. On Product Fidelity, the primary focus of this work, FidPoster achieves 4.838, exceeding GPT-Image-2's score of 4.746.

It is worth emphasizing that this work primarily focuses on accurately preserving fine-grained product information, including packaging text, logos, textures, and colors, rather than comprehensively optimizing the aesthetic quality and creative design of posters. We also observe that GPT-Image-2 occasionally regenerates product packaging text or local patterns into content that differs from the reference product in some generated results. Representative examples are provided in Appendix E.7 (Figure 13). These results demonstrate that FidPoster achieves perceptual quality close to that of a strong closed-source model while obtaining a higher score on the fine-grained product fidelity task targeted by this work, highlighting the practical value of the proposed fidelity-specific mechanisms.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer 2A3x,

Thank you once again for your valuable comments. As the discussion phase is approaching its end, we would like to kindly ask whether our clarifications and additional experiments have sufficiently addressed your concerns. We would be happy to provide further clarification if needed.

Thank you for your time and consideration.

Reviewer_DJemOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses fine-grained detail fidelity in product poster generation, focusing on the preservation of packaging text, textures, and colors. FidPoster fine-tunes Qwen-Image-Edit-2509 with Pixel-Space Fidelity Activation (PFA), which adds time-weighted appearance, edge, and structural losses to flow matching by comparing decoded clean-image estimates with target posters within annotated product regions. A jointly trained reference-aware localizer identifies the product region during inference. At selected sampling steps, Energy-Guided Velocity Correction (EVC) adjusts the sampling velocity using gradients of a bidirectional energy that compares RGB, edge, and high-frequency descriptors between the generated product region and the reference image. The authors curate 14,067 training pairs in FidPoster14K, using products extracted from target posters as references and providing product masks and design instructions. Evaluation on FidPoster-Bench uses 800 held-out product-poster pairs and 200 independently sourced product images. The paper reports improved product fidelity in automatic and blinded human evaluations.

Critical strengths

I acknowledge the empirical strengths below, but I do not consider them sufficient to support acceptance because the paper has not yet established a substantial contribution beyond existing product-preservation approaches.

1. The experiments show fidelity improvements within the evaluated generative setting. Table 4 reports improvements over standard SFT in Product Text Score (82.09 to 84.92) and SSIM (0.7936 to 0.8561) with PFA. On 200 independently sourced products, the blinded human study also reports gains for the complete FidPoster system over Qwen-Image-Edit-2509 in product fidelity (+0.195) and commercial usability (+0.340) on a 1–5 scale, with both paired 95% confidence intervals above zero. These findings support the effectiveness of PFA within the tested backbone and indicate perceived benefits of the complete system on independent inputs.

Critical weaknesses

1. The added value of the proposed fidelity-preservation mechanism over inpainting-based methods is not established. The authors should first clarify whether changes to the reference product's 3D viewpoint are within the intended input-output task and count as valid outputs. If they are, the paper needs to explain whether and how its training and evaluation data include such changes, and demonstrate that its fidelity mechanism remains effective under them. Section 3.1 extracts reference products from their own target posters, suggesting that the resulting training pairs primarily supervise preservation of the product's existing viewpoint. The Independent Reference study tests independently sourced products but does not isolate viewpoint variation. Changes in 3D viewpoint alter projected texture and edge geometry and which surfaces are visible, so the RGB, edge, and high-frequency descriptors used by EVC are not inherently viewpoint-invariant and require dedicated validation under such changes.

If the intended task preserves the reference viewpoint, the authors should explicitly acknowledge this restriction in the task definition and limitations. The paper should discuss the benefits and limitations of its fidelity-preservation mechanism relative to inpainting-based methods and include results against at least one representative inpainting-based method, such as PosterMaker. The authors do not compare against PosterMaker, citing its requirement for product masks and text positions. However, PosterMaker's empty-text support enables a focused subject-preservation comparison, and masks can be obtained using existing segmentation or matting models. A strictly controlled comparison within the same backbone is not required. As an optional further analysis, a comparison using an inpainting-based model such as FLUX-Fill under reasonably matched settings could help interpret the differences. If the proposed mechanism provides comparable or worse fidelity under the same viewpoint restriction, I do not consider avoiding an externally supplied mask alone a sufficiently substantial benefit.

Questions for authors
  1. Clarification of the task and data. Does the task require preserving the reference product's 3D viewpoint, or are viewpoint changes considered valid outputs? Given the extraction procedure in Section 3.1, please clarify whether the training pairs contain viewpoint changes and whether the evaluation explicitly tests them.

  2. Request for evidence supporting the contribution. If viewpoint changes are in scope, what evidence shows that the proposed fidelity mechanism remains effective under such changes? If the reference viewpoint is preserved, please state this limitation, discuss the relationship to inpainting-based methods, and provide results against at least one representative inpainting-based method, such as PosterMaker, with relevant differences in settings reported. What substantial fidelity or practical advantage do these results establish beyond avoiding an externally supplied mask?

  3. Clarification of EVC's substantive contribution. Relative to backward clean-prediction guidance (Bansal et al., Universal Guidance for Diffusion Models, 2023, Sec. 3.2, link), what substantive technical or practical benefit is introduced by EVC's product localization and correspondence energy? Which available results support attributing this benefit to EVC beyond the fidelity already learned through PFA?

4 comments
AuthorsOpenReview ↗

Response to reviewer DJem

Thank you for your supportive review and suggestions. Below we respond to the comments in Weaknesses (W) and Questions (Q). The newly added model comparison experiments in the revised manuscript follow the same evaluators as those used in Table 3 of the main paper to ensure comparability across different methods. Our responses are as follows.

W1 & Q1: Task Scope and Data Coverage Regarding Product Viewpoint Changes

Our task is to generate a complete product poster from a reference product image and a design instruction. Reasonable changes in product orientation and perspective are considered valid outputs, but accurate reconstruction under large 3D viewpoint changes is not the primary objective of this work. We focus on preserving fine-grained visual features, including packaging text, logos, colors, textures, and local structures, that remain correspondable between the reference product and the generated poster.

Regarding the training data, the reference products in FidPoster14K are directly extracted from their corresponding target posters. Therefore, training pairs maintain consistent product viewpoints. The dataset is primarily designed to learn complete poster generation and product detail preservation, rather than to provide explicit supervision for cross-view 3D reconstruction. Regarding the evaluation data, the 800 Fine-grained Fidelity samples are primarily used for controlled product fidelity evaluation with reliable visual correspondences, while the 200 Independent Reference samples use independently sourced product images and manually constructed design instructions, with some generated results exhibiting viewpoint changes.

To further analyze this aspect, we provide qualitative examples in Appendix E.5 (Figure 11) showing changes in product orientation and perspective. These examples demonstrate that FidPoster can preserve recognizable product appearance features under certain viewpoint adjustments.

W1 & Q2: Added Value over Inpainting-Based Methods and Experimental Validation

Inpainting-based methods use reference images and masks to constrain generation, but they do not necessarily preserve all fine-grained product details. For example, PosterMaker conditions generation on the product image and mask rather than directly copying the original product pixels, so packaging text, logos, and local textures may still be distorted.

This is closely related to the two challenges identified in our Introduction. First, latent-space representations do not guarantee faithful decoded details. As shown in Table 1, VAE reconstruction alone reduces the average SSIM from 1.0000 to 0.8043, with an HF/LF error ratio of 3.00. Although this experiment does not directly measure PosterMaker, it demonstrates the risk of fine-grained detail loss in latent reconstruction. Second, reference conditioning alone lacks explicit fidelity feedback to detect and correct product-specific discrepancies during generation. FidPoster addresses these challenges through PFA, which directly supervises decoded product details during training, and EVC, which provides reference-aware fidelity correction during inference.

Moreover, FidPoster preserves flexibility in product presentation and scene integration. As illustrated in Appendix E.6 (Figure 12), it supports natural interactions between products and functional visual effects, such as airflow and water, while plausibly completing partially cropped or occluded product regions. Unlike approaches that strictly preserve fixed product pixels, FidPoster allows the product and surrounding scene to be generated jointly while maintaining fine-grained fidelity.

To further evaluate its practical advantages, we additionally compare FidPoster with PosterMaker, providing the required product masks according to its inference protocol. Following the reviewer's suggestion, PosterMaker is evaluated under an empty-text and simplified-background setting. We conduct a human evaluation on 200 Independent Reference samples:

MethodProduct Fidelity ↑Layout & Composition ↑Visual Quality ↑Commercial Usability ↑Overall ↑
PosterMaker (empty-text)3.9421.0271.2731.2011.861
FidPoster (Ours)4.8384.3264.2444.2354.411

FidPoster improves Product Fidelity by 0.896 points over PosterMaker (4.838 vs. 3.942). Since PosterMaker uses an empty-text and simplified-background setting, its other scores are reported for completeness but are not used to draw conclusions about relative full-poster generation quality.

Overall, these results demonstrate that FidPoster's contribution extends beyond eliminating external masks. By combining decoded pixel-space supervision with reference-aware sampling correction, FidPoster more effectively preserves fine-grained product details while supporting flexible, end-to-end poster generation.

AuthorsOpenReview ↗

Response to reviewer DJem

W1 & Q3: Technical Contributions of EVC Relative to Universal Guidance and Its Independent Benefits

EVC develop reference-aware product localization and bidirectional detail correspondence energy specifically for fine-grained product fidelity in poster generation. These designs enable the guidance signal to directly address local detail discrepancies between the generated product and the reference image.

Specifically, EVC incorporates two key designs.

(1) Reference-aware Product Localization. The position, scale, and layout of the product may change throughout the poster generation process, and applying global image constraints directly can interfere with the background and other design elements. To address this issue, we use the reference-aware localizer trained during the PFA stage to identify the current generated product region during guidance and restrict the fidelity gradient to this region, thereby focusing sampling correction on the product itself.

(2) Bidirectional Detail Correspondence Energy. Instead of relying on strict pixel-wise alignment between the reference image and the generated product, EVC establishes local correspondences using RGB, edge, and high-frequency features. Specifically, generated-to-reference matching suppresses incorrect or spurious details, while reverse reference-to-generated matching constrains the omission of important reference details. In this way, EVC provides reference-aware fidelity feedback for packaging text, logos, local structures, and textures even when the product undergoes changes in position and scale, and uses this feedback to correct the Flow Matching sampling velocity.

Regarding viewpoint changes, bidirectional matching can provide useful fidelity guidance under moderate viewpoint variations when product details remain locally correspondable, without requiring strict pixel-wise alignment. The qualitative examples in Appendix E.5 (Figure 11) illustrate that FidPoster can preserve recognizable product details under certain changes in product orientation and perspective.

To evaluate the independent contribution of EVC beyond PFA, we compare the results with EVC disabled and enabled under the same backbone and training weights. On the Independent Reference evaluation set, we additionally conduct human evaluations of Standard SFT and PFA-only, with the following results:

MethodProduct Fidelity ↑Layout & Composition ↑Visual Quality ↑Commercial Usability ↑Overall ↑
Baseline4.6434.1114.0093.8954.165
+ Standard SFT4.6774.2804.1804.1324.317
+ PFA (w/o EVC)4.7704.3064.2024.2124.373
FidPoster (PFA + EVC)4.8384.3264.2444.2354.411

On Product Fidelity, the primary focus of this work, Standard SFT achieves 4.677, which increases to 4.770 with PFA and further improves to 4.838 (+0.068) when EVC is enabled. Meanwhile, Visual Quality increases from 4.202 for PFA-only to 4.244, and Overall increases from 4.373 to 4.411. These results demonstrate that EVC provides additional fidelity improvements during the generation of individual products, even after the model has learned product detail preservation through PFA.

Thank you for clarifying the training data and the intended scope of viewpoint changes, and for providing the additional comparisons and human evaluations.

Your clarification confirms that the training pairs strictly preserve the product viewpoint. The supervised task is therefore viewpoint-preserving product generation, consistent with the preservation setting of inpainting-based methods. The ability to change viewpoints at inference is inherited from the pretrained backbone, while training exclusively on these pairs narrows its capabilities toward viewpoint preservation. I therefore assess the contribution primarily within this setting.

Within this setting, I now recognize the practical value of the proposed fidelity mechanisms in light of the current trend among SOTA image editing models: their standard editing interfaces rely on reference images and textual instructions without native mask inputs to enable inpainting-based compositing that preserves the original product pixels. Specialized inpainting-based poster models cannot simply be transferred to these backbones. PFA and EVC offer a way to strengthen product fidelity within this increasingly prevalent editing framework, allowing product poster generation to benefit from SOTA editing backbones without switching to a specialized inpainting-based poster model.

The additional human evaluations also support improvements beyond standard SFT: Product Fidelity increases from 4.677 with Standard SFT to 4.770 with PFA and 4.838 with PFA + EVC on independently sourced products. These results strengthen the evidence for the contribution of the proposed fidelity mechanisms. The PosterMaker and SimplePoster comparisons should still be interpreted with their different backbones and native prompt/input formats in mind.

I encourage the authors to make the viewpoint-preserving training task and the above practical motivation clearer in the manuscript. Based on the clarifications and additional evidence, I am raising my rating to weak acceptance.

AuthorsOpenReview ↗

Thanks for the reply

We sincerely thank you for your thoughtful follow-up and constructive feedback throughout the review process. We are pleased that our methodological contributions and additional experimental evidence have been recognized, and we greatly appreciate your positive reassessment and increased rating. We will further clarify the training setting and its practical motivation in the revised manuscript, as suggested.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Response to reviewer MnQZ

W1 & Q1: Fairness of Ablation Studies and Contributions of PFA and EVC

We have clarified in the revised manuscript that EVC is disabled in all Table 4 ablations, and that the Improvement row in Table 2 measures gains over the best zero-shot baseline for each metric. To further isolate the contributions of PFA and EVC beyond standard SFT, we additionally evaluate Standard SFT and PFA-only on the 200 Independent Reference samples using the same human evaluation protocol.

MethodProduct Fidelity ↑Layout & Composition ↑Visual Quality ↑Commercial Usability ↑Overall ↑
Baseline4.6434.1114.0093.8954.165
+ Standard SFT4.6774.2804.1804.1324.317
+ PFA (w/o EVC)4.7704.3064.2024.2124.373
FidPoster (Ours)4.8384.3264.2444.2354.411

On Product Fidelity, Standard SFT achieves 4.677, which improves to 4.770 (+0.093) with PFA and 4.838 (+0.068) with EVC, yielding a total gain of 0.161 beyond standard SFT. Importantly, these gains persist on independently sourced products without paired target posters, indicating that the improvements are not restricted to the paired evaluation setting.

These results distinguish the contributions of the two stages: PFA improves fine-grained product detail preservation through decoded pixel-space supervision, while EVC provides additional reference-aware correction during inference. Together, they demonstrate that FidPoster's fidelity improvements cannot be attributed solely to standard SFT.

Q2: EVC Implementation Details and Hyperparameter Sensitivity

We have further clarified the relationship between the equations and the implementation details of EVC. Equation (6) in the main paper and Equation (17) in the appendix define the same bidirectional product fidelity energy. The only difference is that the appendix explicitly specifies the summation range and further explains the computation details of each energy component. The relationship between the energy gradient and the sampling velocity correction is fully derived in Appendix D.3.

To improve the reproducibility of our method, we have explicitly specified the following hyperparameters in the revised manuscript:

The specific hyperparameters are as follows:

ParameterValue
Reference Saliency Weights ωj\omega_j∑jωj=1\sum_j \omega_j = 1
Reference Coverage Weight λr\lambda_r0.8
Time Lower Bound tmin⁡t_{\min}0.05
Soft-Min Temperature τ\tau0.07
Localizer Sharpness β\beta5.0

To verify the stability of EVC, we further conducted sensitivity experiments on Guidance Scale and Guidance Interval using FidPoster-Bench. With the backbone model, test samples, and other inference configurations kept unchanged, we obtained the following results.

Guidance Scale (η\eta)Guidance IntervalP-Text ↑SSIM ↑GMSD ↓Color ↓HF ↓
0.4040%–90%85.010.86840.10433.4100.9798
0.4055%–90%85.120.86840.10463.4050.9952
0.4065%–95%85.340.86780.10503.4280.9709
0.8065%–95%85.120.86590.10553.4910.9771
0.6040%–90%84.750.86870.10413.3830.9656
0.6065%–95%85.300.86670.10523.4700.9750
0.6055%–90%85.490.86300.10663.3790.9709

The results demonstrate that EVC maintains relatively stable product detail fidelity across different guidance strengths and application intervals. Different parameter configurations involve certain trade-offs across individual metrics. These results indicate that the effectiveness of EVC is not limited to a particular set of hyperparameters, as different configurations can all maintain good product detail fidelity. We have included the relevant parameters and sensitivity experiments in the Implementation Details and Appendix E.10 of the revised manuscript to further improve the transparency and reproducibility of our method.

AuthorsOpenReview ↗

Response to reviewer MnQZ

W2 & Q3: Benchmark Construction and Design Prompt Dependency

(1) Evaluation objectives and dataset construction. Our work focuses on Fine-grained Product Detail Fidelity, rather than autonomous layout planning or creative design. Accordingly, P-Text, SSIM, GMSD, Color Difference, and HF primarily assess product detail preservation, while Poster Text Score and FID serve as auxiliary metrics for promotional-text accuracy and overall poster quality.

FidPoster-Bench contains two complementary evaluation settings. The 800 Fine-grained Fidelity samples use Gemini 2.5 Pro to extract design instructions from target posters, covering promotional copy, product positions, typography, colors, backgrounds, and styles, followed by manual verification. Reference products are extracted from the corresponding target posters, enabling controlled automatic fidelity evaluation.

The 200 Independent Reference samples use independently sourced product images and manually constructed design instructions without target posters. Since viewpoint variations may affect geometrically aligned metrics such as SSIM and GMSD, we adopt human evaluation to assess product fidelity. FidPoster achieves 4.838, compared with 4.643 for the baseline, demonstrating improvements on independent products. Thus, the two settings complement each other through controlled quantitative evaluation and independent-product generalization assessment.

(2) Influence of design instructions on product fidelity. Design instructions specify layout, typography, and visual style, but do not systematically provide complete packaging text, logos, or fine-grained textures. Therefore, preserving these details still requires accurate visual information from the reference image.

To examine prompt dependency, we independently remove color, promotional copy, or position descriptions on the 800 Fine-grained Fidelity samples, keeping other settings unchanged (Appendix E.9).

Prompt SettingP-Text ↑SSIM ↑GMSD ↓Color ↓HF ↓
w/o Color84.270.86830.10353.4480.9628
w/o Copy83.170.85490.10943.7531.0070
w/o Position84.870.86130.10723.6940.9824
Full Prompt (FidPoster)85.490.86300.10663.3790.9709

The results show that FidPoster maintains relatively high product fidelity when individual design constraints are removed, although different prompt components affect individual metrics. Notably, removing color descriptions slightly improves SSIM, GMSD, and HF.

Q4: Additional Poster Generation Baselines and Automatic Fidelity Evaluation

(1) Additional comparison with SimplePoster. As we have not identified publicly reproducible official weights and inference code for DreamPoster, we additionally evaluate SimplePoster on FidPoster-Bench. SimplePoster requires product-region masks for its inpainting-based generation, which we provide according to its inference protocol, whereas FidPoster requires no externally provided mask.

We evaluate both methods on the same 200 Independent Reference samples, following the five-rater blind evaluation protocol of Table 3.

MethodProduct FidelityLayout & CompositionVisual QualityCommercial UsabilityOverall
SimplePoster4.0933.8833.2902.8053.518
FidPoster (Ours)4.8384.3264.2444.2354.411

FidPoster improves Product Fidelity by 0.745 and Overall by 0.893, demonstrating stronger fine-grained product preservation while maintaining complete poster generation quality.

(2) Automatic fidelity evaluation on Independent Reference. The Fine-grained Fidelity setting supports automatic evaluation using paired target posters, whereas Independent Reference contains independently sourced products without target posters. Geometrically aligned metrics such as SSIM and GMSD may penalize reasonable viewpoint changes, motivating our use of human evaluation.

To provide additional automatic evidence, we evaluate P-Text on the first 100 Independent Reference samples. Unlike pixel-alignment metrics, P-Text measures packaging text consistency between generated and reference products without requiring paired target posters.

MethodP-Text ↑
BAGEL47.68
FLUX.276.44
IC-Custom63.61
OmniGen225.88
PosterOmni54.09
Qwen-Image-Edit-250977.17
Step1X-Edit85.67
FidPoster (Ours)87.02

FidPoster achieves the highest P-Text score of 87.02, providing additional automatic evidence of packaging text preservation on independently sourced products.

AuthorsOpenReview ↗

Follow-Up on Our Rebuttal

Dear Reviewers,

We sincerely thank all reviewers for your constructive feedback. We have carefully addressed all the Weaknesses and Questions raised by each reviewer in our detailed point-by-point responses. We have also revised the manuscript and conducted additional experiments.

As the discussion phase is approaching its end, we would greatly appreciate it if you could let us know whether our clarifications and additional experiments have adequately addressed your concerns. We would be happy to provide further clarification or discussion if needed.

Thank you again for your valuable time and thoughtful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.