Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Training-Free Constrained Generation With Stable Diffusion Models
Authors: Stefano Zampini, Jacob K Christopher, Luca Oneto, Davide Anguita, Ferdinando Fioretto
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of this approach is demonstrated through material design experiments requiring adherence to precise morphometric properties, challenging inverse design tasks involving the generation of materials inducing specific stress-strain responses, and copyright-constrained content generation tasks. All code has been released at https://github.com/RAISELab-at UVA/Constrained-Stable-Diffusion. |
| Researcher Affiliation | Academia | Stefano Zampini Polytechnic of Turin EMAIL Jacob K. Christopher University of Virginia EMAIL Luca Oneto University of Genoa EMAIL Davide Anguita University of Genoa EMAIL Ferdinando Fioretto University of Virginia EMAIL |
| Pseudocode | Yes | Algorithm 1: Sampler with Constraint Correction Input: δ (violation tolerance), lr (learning rate) Define prox_objective(xi t): violation g(xi t); distance 1 2λ xi t x0 t 2 2; return violation + distance; for t T to 0 do // Sampling steps (omitted). while g(D(zi t)) δ do g zi tprox_objective(D(zi t)); zi+1 t zi t (g lr); i i + 1; Output: D z0 |
| Open Source Code | Yes | All code has been released at https://github.com/RAISELab-at UVA/Constrained-Stable-Diffusion. |
| Open Datasets | Yes | For this experiment we obtain the dataset used by [15, 39]. From [38], we obtain a dataset of periodic stochastic cellular structures subjected to large-strain compression. |
| Dataset Splits | No | The paper mentions using specific datasets, for example, 'a balanced dataset of 8,000 images is constructed to fine-tune the classifier and diffusion models' in Section 6.3, but does not explicitly provide training/test/validation split percentages, sample counts, or detailed splitting methodologies for any of the datasets used. |
| Hardware Specification | Yes | Table 1: Experiment runtimes by sampling type and hardware. Microstructure Conditional Sampling 10 s Nvidia A100-SXM4-80GB Microstructure Constrained Sampling 50 s Nvidia A100-SXM4-80GB Metamaterials Conditional Sampling 5 s Nvidia A100-SXM4-80GB Metamaterials Single Simulation 30 s Intel Core i7-8550U CPU Copyright Conditional Sampling 10 s Nvidia A100-SXM4-80GB Copyright Constrained Sampling 65 s Nvidia A100-SXM4-80GB |
| Software Dependencies | No | The paper mentions software components like 'Abaqus [40]' (a simulator) and the use of a 'Res Net50 backbone' for a classifier, but it does not provide specific version numbers for these or any other key software libraries, frameworks, or programming languages used in the experiments. |
| Experiment Setup | Yes | Algorithm 1: Sampler with Constraint Correction Input: δ (violation tolerance), lr (learning rate) ... For this implementation, we set the number of perturbed samples M = 10, finding this provides strong enough approximations of the gradients to converge to feasible solutions. ... As shown in Figure 3, with five DPO steps our method provides a 4.6x improvement ... The architecture of the classifier consists of a Res Net50 backbone, which is followed by two fully connected layers. These layers serve to progressively reduce the dimensionality of the feature map, first from 2048 to 512 and then from 512 to a single scalar feature, which represents the output of the classifier. A Sigmoid activation function is then applied to this final feature to estimate the probability that the input sample belongs to either the Mickey Mouse or Jerry class. |