acceptodds
Under review as a conference paper at ICLR 2027

Jewel-T2I: Learning Jewelry Generation from Expert Seed Data

Abstract

T2I generation can support jewelry customization by visualizing design requirements before physical prototyping, but general models often confuse craft structures or fail to preserve the requested motifs and materials. We present Jewel-T2I, which constructs training data from expert jewelry assets for domain adaptation. Its data pipeline, JewelGen, expands 500 seed assets through semantic and viewpoint editing, while visual craft exemplars guide the synthesis of plain, filigree, and enamel designs. MLLM screening and designer verification select 20K image-text pairs for adapting the diffusion model. The resulting generator is combined with a prompt model trained on a separate curated jewelry corpus, allowing users to specify design requirements without supplying reference images at inference. We also introduce JewelBench, a benchmark of 2K entries that evaluates AF, CF, and AQ across the three crafts. With identical prompts, the adapted generator improves over Qwen-Image by 6.3, 8.6, and 5.0 points on these criteria, respectively. The complete system achieves the highest mean scores among the compared systems across all three crafts under both automated and human evaluation. These results show that the constructed training data improves the ability of T2I models to follow jewelry design requirements on JewelBench.

Primary area:
generative models
Keywords:
Text-to-Image, Benchmark, Dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_crbyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

JEWEL-T2I expands 500 expert jewelry assets through semantic editing, viewpoint editing, and synthesis conditioned on craft exemplars. Gemini screening and designer verification select 20,000 image–text pairs used to LoRA-adapt Qwen-Image into JewelDiffusion.

A separate Qwen2.5-1.5B model is adapted on 10,000 curated text examples to generate jewelry descriptions. The complete system is evaluated on JewelBench, containing 2,000 entries across plain, filigree, and enamel designs. Evaluation measures attribute fidelity, compositional fidelity, and aesthetic quality through automated and human assessment.

Critical strengths

The author raised a concrete application need that Jewelry generation requires attention to craft appearance and motif relationships that broad image-quality scores may overlook, which is interesting to me.

And a well-documented, independently split jewelry benchmark could benefit research in specialized visual generation.

Critical weaknesses

My major concern is that synthetic-sample selection followed by image-model fine-tuning is established, and the paper does not introduce any new learning objective or mechanism for satisfying craft constraints.

The recipe could be reused in other design domains, but its general applicability largely follows from existing domain-adaptation techniques. Cross-domain experiments are not mandatory; a clearly demonstrated new insight within jewelry would also suffice. That insight is currently underdeveloped.

Questions for authors
  1. How much source-family overlap occurs across the current split, and what results hold when all descendants of each original seed remain in one partition?
  2. What improvement does the full pipeline provide over seed-only adaptation and equal-sized synthetic datasets generated without craft exemplars or expert filtering?
  3. Which exact evaluator model, prompt, and code produced the reported results? Can you provide the matching implementation and anonymized human ratings?
  4. How does JewelExpert compare with unadapted Qwen2.5-1.5B, and do the generator gains persist when craft correctness is evaluated separately from material, lighting, and viewpoint?
Reviewer_ZzXAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents JEWEL-T2I, a jewelry image generation method from text as input for addressing domain-specific data scarcity, modeling specialization and fine-grained evaluation. To address these issues, the method proposes three contributions, JEWELGEN, a data construction pipeline; JEWELCRAFT, which combines a prompt model trained on curated jewelry text with a diffusion model adapted on generated image-text pairs; JEWELBENCH, a dataset for evaluating attribute fidelity, compositional fidelity, and aesthetic quality.

Critical strengths
  • This work is well-motivated for an under-explored domain. The studied problem is important.
  • It provides a practical data-model-evaluation loop with consistent improvements over strong baselines.
  • JEWELGEN starts from 500 seed samples and designs a concrete pipeline for data filtering and generation, which provides a practical way for generating low-resource data. It also includes expert review.
  • The experiments evaluate both the full system and its components
Critical weaknesses
  • The core models are based on existing Qwen models adapted with LoRA, so the proposed method has limited contributions to network architecture design.
  • It is not clear whether this model has a strong generalization ability, since the data were split randomly after augmentation, and thus variants of the same source design may appear in both training and evaluation.
  • The experiments do not provide a clear diversity metric, so it is not clear about this model's design diversity and out-of-distribution performance.
Questions for authors
  • The proposed method's output is visual images. How to make sure that the designed jewelry can be fabricated?
  • The benchmark covers plain, filigree, and enamel designs but excludes gemstone settings and other techniques, so it raises a question about the scope and applicability of this method.
Reviewer_4tG6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed text-to-image generation for jewelry design. The authors propose JewelGen, which starts from 500 expert seed assets and generates candidate images through semantic editing, viewpoint editing, and visual craft exemplars. The candidates are then filtered by Gemini and reviewed by jewelry designers to construct training data. The authors fine-tune Qwen-Image on 20K image-text pairs and separately fine-tune Qwen2.5-1.5B on 10K jewelry text samples to generate design prompts. They also introduce JewelBench, which contains 2K jewelry samples covering plain, filigree, and enamel designs, and evaluates them using attribute fidelity (AF), compositional fidelity (CF), and aesthetic quality (AQ).

Critical strengths
  1. General T2I models may have a gap between their generated jewelry designs and human requirements. The authors make an attempt to adapt generation models to the jewelry domain.
Critical weaknesses
  1. Source-level overlap between the training and test sets. Different edited versions of the same seed asset may appear in both the training set and JewelBench. The model may therefore learn the contours, compositions, and textures of specific jewelry items rather than generalizable craft and design relationships.
  2. The method contribution mainly combines existing components. In particular, the authors do not provide sufficient ablations for many components. JewelExpert is also not compared with an unfine-tuned Qwen model. Moreover, the method does not introduce a particularly targeted design or modeling mechanism for jewelry beyond data construction.
  3. Some conclusions are overstated. Parts of the paper suggest that designer verification ensures the manufacturing plausibility of the generated results. However, the limitations section acknowledges that generated images alone cannot establish physical manufacturability.
  4. The automatic evaluation relies entirely on Gemini-based scoring. AF, CF, and AQ are all assigned by Gemini according to natural-language criteria, without explicit computation rules, objective annotations, or independent validation. Therefore, they are not sufficiently rigorous or reproducible benchmark metrics.
Questions for authors
  1. Can the authors provide results on a source-disjoint JewelBench test set?
  2. Compare their method and give the results with more recent MLLMs, such as GPT-6 and GPT Image 2, as well as more recent T2I models? The current baselines appear relatively outdated.
  3. Does JewelExpert provide advantages over general-purpose models such as GPT-6 or Claude Opus 5.5 for rewriting jewelry design requirements?
  4. Can the authors provide ablations for the different components and clarify the contribution of each part?
Reviewer_TGYTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a pipeline for adapting text-to-image generation to jewelry design. JEWELGEN expands 500 seed assets through semantic and viewpoint editing, synthesis conditioned on visual craft exemplars, and automated screening followed by designer verification. The resulting 20K image-text pairs are used to adapt Qwen-Image, while a separate 10K text corpus supports a prompt-generation model. The paper also introduces JEWELBENCH, comprising 2K entries across plain, filigree, and enamel jewelry. Experiments report improvements in attribute fidelity, compositional fidelity, and aesthetic quality through fixed-component comparisons and automated and human evaluation.

Critical strengths
  1. The fixed-prompt comparison provides useful evidence for the benefit of domain adaptation. With identical prompts and matching inference settings, JEWELDIFFUSION improves over Qwen-Image by 6.3, 8.6, and 5.0 points in AF, CF, and AQ, respectively. Holding these factors fixed makes this comparison particularly informative: it supports an improvement attributable to adapting the image generator on the constructed dataset under the reported evaluation protocol.
  2. The evaluation addresses requirements specific to jewelry design and includes substantial human assessment. The rubric distinguishes craft and material accuracy from motif composition and visual appeal. Five designers and five users independently assess each system-comparison output, with model names hidden and image order randomized. The consistent ranking of JEWELCRAFT across automated and human assessment provides complementary support for the reported system-level gains.
Critical weaknesses
  1. Generalization beyond the source designs remains unresolved. Training and benchmark images are partitioned after augmentation without grouping by source seed (§3.1). As acknowledged in §6, variants of the same design may therefore occur in both sets. This limits the extent to which the reported gains establish performance on unseen designs, an important consideration when assessing the value of both the training pipeline and the benchmark.
  2. The benefit of the distinctive data-construction strategy is not sufficiently isolated. Table 4 demonstrates the benefit of adaptation using the complete dataset. However, expert-guided data construction is a central contribution, and the experiments do not establish the contribution of visual craft exemplars relative to a simpler synthetic-data pipeline with comparable training data and compute. Section 6 acknowledges the absence of construction-step ablations.
Questions for authors
  1. Could you provide results with source designs held out before augmentation and excluded from training-time exemplar selection? This would clarify whether the gains persist beyond related variants of the seed designs and directly address the limitation acknowledged in §6.
  2. Could you compare the full pipeline with a version that omits visual craft exemplars, using the same seed pool, filtering procedure, training-data size, and model-training settings? This comparison would help establish the contribution of a central JEWELGEN component beyond the overall adaptation benefit reported in Table 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.