acceptodds
Under review as a conference paper at ICLR 2027

StructRDM: Prototype-Anchored Residual Diffusion for Titanium Alloy Microstructure Generation

Abstract

Titanium alloy microstructures are jointly influenced by alloy composition, processing parameters and mechanical properties. In limited-data regimes, however, these low-dimensional conditions underdetermine the spatial arrangement of phase regions and the geometry of phase boundaries in high-dimensional microstructure images, making end-to-end conditional generation prone to fragmented phase regions and unstable boundaries. To address these issues, we propose StructRDM, a reliability-aware, prototype-anchored residual diffusion framework. StructRDM follows a “structure first, texture second” strategy. It first predicts structural statistics from multi-source material conditions and estimates their reliability through conformal calibration. It then retrieves type-consistent structural prototypes through reliability-weighted matching and generates the target phase structure using latent residual diffusion. Finally, the generated phase-structure map provides spatial guidance for microstructural texture synthesis. Compared with the direct condition-to-image diffusion baseline, StructRDM improves F1 by 26.9% while reducing Gradient W1 and Autocorrelation MAE by 21.0% and 35.4%, respectively, indicating broader distributional coverage and improved spatial-statistical consistency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Titanium alloy microstructures, conditional generation, diffusion models, prototype anchoring

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 3 reviews, above 27% of the others.
Reviewer_tf7nOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

StructRDM studies titanium-alloy microstructure generation with limited paired data. It predicts structural statistics and their uncertainty from material conditions, retrieves a real structural prototype, and learns a residual correction before synthesizing texture. The central idea is to use real spatial layouts to simplify generation of complex microstructures from global material conditions.

Critical strengths

Real prototypes provide useful spatial layouts and improve morphology with limited data. Global composition and processing conditions do not directly specify complex regional arrangements, so a prototype gives generation a useful structural starting point. Removing prototypes causes substantially more fragmentation in Table 3, and the final-image spatial statistics in Table 4 support improved morphology. This is the most convincing benefit. The lower distributional precision in Table 4 also indicates that the gains concern spatial organization and coverage.

Critical weaknesses
  1. Following predicted statistics does not sufficiently establish effective material conditioning. Material conditions are first compressed into structural statistics and uncertainty estimates, which then guide retrieval and generation. SCE measures agreement with these predicted statistics. Table 1 evaluates the predictor, but these results and SCE do not establish that changes in composition and processing within a microstructure type are appropriately reflected in the generated structures. This affects the central value of the method as a material-conditioned generator.

  2. The ablation does not separate the benefits of real prototypes and learned corrections. Removing prototypes in Table 3 also removes the prototype-residual representation, so the substantial degradation establishes the value of the combined design. The method additionally encourages preservation of the retrieved structure. Without comparing against direct prototype reuse with the same texture-generation stage, it remains unclear whether residual diffusion learns useful corrections for the target conditions.

Questions for authors
  1. Within a fixed microstructure type, what evidence shows an appropriate response to changes in composition and processing, assessed independently of the model's own predicted statistics?

  2. How does directly passing the retrieved prototype to the same texture-generation stage compare with the full method in morphology and agreement with the target conditions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer tf7n: Within-Type Conditioning and Prototype Reuse

1. Within-type material-condition correspondence. We thank the reviewer for pointing out that SCE measures agreement with model-predicted statistics and cannot independently validate the role of material conditions. We added a within-type A/B matching test. From the fixed test set, we select pairs with the same microstructure type, different sources, and different compositions or recorded processing conditions. For each pair, we compare the sum of structural-statistic errors under the correct assignment of generated images to their paired real images with the sum under the swapped assignment; a lower error under the correct assignment counts as a correct match. The six structural statistics are extracted from the paired real images and final generated images. Scoring does not use the model-predicted statistical targets.

Across 170 test images and four microstructure types, StructRDM achieves a class-averaged correct-matching rate of 61.5% (95% CI: 52.8%–71.7%, bootstrapped over 97 source groups). Randomly permuting generated-image assignments within each type gives approximately 50.0%; a 2,000-permutation test yields p=0.001p=0.001. Thus, even within a fixed type, the generated images retain sample-specific structural information associated with different input material conditions. This evaluation complements the predictor results in Table 1 and SCE by using paired real images as an independent reference.

2. Prototype reuse and residual refinement. We thank the reviewer for noting that the original “w/o Prototype Anchoring” ablation removes both real prototypes and the residual representation, so it cannot isolate the contribution of residual refinement. We therefore added a direct Rank-1 prototype reuse control: without residual refinement, the retrieved phase map is passed to the same stage-two generator as in the full model. Both use the submitted retrieval setup, the same 170 test samples, stage-two weights, and sampling randomness. We also verified that the full-model outputs match the original Table 4 images one by one.

MetricRetrieved Prototype OnlyFull StructRDM
Phase-map SCE ↓0.69490.6853
Phase-map Small Component Error ↓0.01980.1058
Final-image MAE ↓0.30420.2829
Final-image MMD² ↓17.76127.0181
Final-image Gradient W1 ↓0.11180.1002
Final-image F1 ↑0.63200.5212

The results show that retrieved prototypes provide the main morphological prior. Residual refinement slightly reduces first-stage SCE and improves final-image MAE, MMD², and gradient-distribution matching. However, it does not improve all morphology metrics; final-image F1 and Condition-SCE, measured against predicted statistics, also worsen. We therefore characterize residual diffusion as a refinement that trades off image-distribution matching against structural coverage, rather than claiming uniform gains in morphology or material-condition adherence.

Thanks for adding the within-type matching test and the matched prototype-reuse control. However, my remaining concern still lies in the contribution of residual refinement. In the paper it shows improvements in image MAE, MMD and gradient statistics, but as phase fragmentation increases, the final-image F1 and condition agreement get worse, and the within-type matching test does not establish an additional benefit over prototype reuse. Thus, the control does not fully support residual generation as a way to improve microstructure and conditional fidelity. This leaves the central added contribution insufficiently demonstrated, so I maintain my original score.

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

The paper proposes StructRDM, a two-stage pipeline for titanium alloy microstructure generation. The approach operates by first invoking a latent residual diffusion model to generate a phase map, which acts as a conditional anchor in the second stage which realizes the microstructure via a secondary diffusion sampling process. The paper targets microstructure generation with higher fidelity to a set of material information by anchoring the first stage to a retrieved prototype. Empirical results are mixed, but report qualitative improvements for select metrics.

Critical strengths
  • Methodological Design: The paper targets an interesting problem, and the two-stage approach is conceptually compelling idea to address this setting. The decoupling of the phase map and the downstream generation provides an useful decomposition of the problem, which could simplify the learning tasks of the individual components.

  • Prototype Retrieval: The retrieval augmentation leveraged by the prototype-anchoring is well motivated, and this seems like an intuitive way to strengthen model performance. Experimental results (Table 3) speak to the significance of this design choice.

  • Dataset Release: The curated dataset could be a nice asset for the community, should the authors intend to release their data. I would consider this a useful contribution of the work.

Critical weaknesses

Evaluation: I have some substantial concerns regarding the evaluation conducted. I will list these one-by-one:

  • The baselines outperform the proposed method on key metrics by a significant margin. Specifically, these metrics, namely MAE\text{MAE}, MMD2\text{MMD}^2 , and Precision\text{Precision} are the metrics which best delineate sample quality. While the authors frame this as "broader distributional coverage", this could also be a sign of sample degradation / inferior sample quality.
  • The metrics which StructRDM performs well on seem to be likely driven by the prototype-anchoring. These metrics (e.g., Recall\text{Recall} and the headline metric F1\text{F1}) do capture diversity, but this is likely "gamed" by access to the training set during generation. The baseline methods cannot exploit this retrieval in the same way, which makes it difficult to compare. It's difficult to assess whether a conditional diffusion model with retrieval would perform any worse on these metrics.
  • To the previous point regarding alternative baselines, it remains unclear if these baselines reasonable. Ikeda-cDDPM is adapted from a different microstructure task, and Stats-DDPM is the authors' own construction. Applying DDPM directly, a conditional DDPM model that generates in a single stage, and a conditional DDPM model with retrieval augmentation for conditioning would be a few ideas I would recommend trying.
  • Metrics are not included to assess whether the downstream samples (across baselines) actually adhere to the conditioning. SCE is included in Table 3, but doesn't seem to be measured in the downstream analysis (Table 4).
  • Confidence intervals should be included, especially given the limited size of the validation set.

Methodological Contribution: While the two-stage pipeline is interesting, the method seems to largely realize its performance due to the retrieval augmentation. This obfuscates whether the methodological contribution is responsible for any performance gains or if a simpler a conditional DDPM model with retrieval augmentation for conditioning could achieve matched performance.

Questions for authors
  • Could you clarify the mismatch between the sampling schedule described in lines 350-356 and the inference parameters reported in Table C.1?
  • Could ground truth phase maps be used in stage two to better understand the downstream effects of error introduced in stage one? If so, has this been explored or evaluated?
Reviewer_RQ94OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a two-stage conditional diffusion model for generating microstructure images of titanium alloy from incomplete physical condition information (e.g. alloy’s composition and mechanical properties) obtained from various sources. The proposed method operates in two stages: first stage produces an intermediate phase structure map by first extracting structural pseudo-labels and morphological descriptors from unannotated micrographs via K-means clustering followed by retrieval of training prototypes that fit the descriptors, and refining them using latent residual diffusion. The second stage generates the texture using a conditional U-net that generates realistic grayscale microstructural textures conditioned upon the phase-structure map from the first stage. Authors provide experimental evaluation on ~1100 alloy samples showcasing reduced phase fragmentation and better statistical consistency compared to vanilla conditional diffusion baselines.

Critical strengths
  • The paper focuses on an interesting problem of significant importance in the material design / discovery domain – the challenge of limited/incomplete heterogenous data is highly real / practical, and such a framework can likely extend to applications beyond titanium alloys.

  • Separating spatial phase-layout generation from local texture synthesis is a logical strategy for limited-data regimes.

  • The evaluation includes statistical metrics like two-point spatial autocorrelation, lineal-path functions, and small-component fragmentation errors that have physical significance associated with them, which makes them more suitable for the given task.

Critical weaknesses
  1. The core premise of Stage 1 is to bridge material conditions and spatial structure via 6D structural statistics. However, in Table 1, the ERT regressor achieves a mean R^2 of only 0.1848. Meaning the signal coming from the material conditions is barely able to predict the structure, breaking the premise.

  2. Prototype retrieval is restricted to candidates of the same ground-truth microstructure class. Thus it seems that the generated phase map is heavily constrained to stay close to a real training sample of the same class. There is no "prototype only" baseline, where the retrieved phase map is passed straight to Stage 2. Similarly, no random same-class retrieval baseline is provided. Without these, it cannot be determined if the residual diffusion and reliability weighting is playing a meaningful role.

  3. As per table 4, the Ikeda-cDDPM surpasses the proposed method on 3 of the 7 metrics: MAE (0.22 vs. 0.28), MMD (5.33 vs. 7.01), and Precision (0.85 vs. 0.55). These inferior values for the proposed method are missing from the abstract (which highlights only the metrics where the proposed method performs better than the baselines). This makes the performance claims of the paper misleading. Moreover, the paper omits comparisons against standard conditional image generation baselines such as ControlNet, or standard LDMs (e.g. Stable Diffusion v1.5) trained with similar auxiliary losses.

  4. It seems that the models are trained using the exact same losses that are used as the metrics in the evaluation. Specifically, in stage 1, the metrics such as area fraction and boundary density also appear as loss terms. Similarly in stage 2, the loss on sobel gradient magnitude is the same as the gradient W1 metric. Also, the MMD is calculated over 159-dimensional feature vectors, and a lot of the components of those feature vectors appear in the loss: gradient orientation histogram, soft edge loss etc. Collectively with the other limitations of the evaluation above, it raises a serious question on the effectiveness of the proposed architectural design, or if the model simply optimized the metric directly.

Questions for authors

N/A

Comments on the paper

1 comment

Thanks authors for detailed response and updates.

While my concerns about presentation (abstract and claims) are addressed, some major concerns about the substance still remain:

  1. The residual diffusion stage gives mixed results. The new prototype-only control shows that most of the structure comes from retrieval. Seems like the rebuttal reports only the metrics where refinement helps: as per Tab. F3, refinement worsens both Macro Dice and small component error by about 5x. This directly contradicts the paper’s claim of reduced phase fragmentation. In Table F.4 (not discussed in rebuttal!), the prototype-only pipeline achieves higher Precision, Recall, F1 etc. F1 is the same metric used to argue superiority over Ikeda-cDDPM, so the necessity of the Stage 1 diffusion model is questionable.

  2. The results are now mixed at best: they are worse on MAE, MMD^2 and Precision.

  3. The argument against ControlNet applies to end-to-end conditioning, but it does not apply to Stage 2, which takes a spatial phase map as input.

Given these remaining concerns, I maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.