Pith. sign in

REVIEW 3 major objections 6 minor 28 references

Stylish and Functional: Guided Interpolation Subject to Physical Constraints

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Guiding diffusion interpolation with a rotational-symmetry constraint yields more realistic, more functional wheel designs.

desk verdict A genuine method extension with plausible FID gains, but the functional-compliance evidence is self-referential and the RA operator as written doesn't enforce n-fold symmetry; worth a revision, not a pass. read the letter →

arxiv 2412.15507 v1 pith:CNRBTVUT submitted 2024-12-20 cs.LG cs.CV

Yan-Ying Chen , Nikos Arechiga , Chenyang Yuan , Matthew Hong , Matt Klenk , Charlene Wu This is my paper · ORCID
classification cs.LGcs.CV
keywords diffusionmodelsinterpolationphysicalconstraintsrotationalsymmetryzero-shotengineeringdesignlatentFID
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper addresses a gap in generative design: blending two reference images of engineered objects often produces visually appealing but physically invalid artifacts. It claims that a zero-shot, training-free intervention in the diffusion process can enforce a physical constraint—here, rotational symmetry of car wheels—while also improving image realism. The proposed FIT framework decodes the latent interpolation at each denoising step, projects the decoded image onto the subspace of rotationally symmetric images via a symmetrizer, and pools the regularized latent back with the original latent using a cosine-similarity-weighted, decaying mix. On a commercial wheel dataset, FIT reduces validation FID from 39.28 to 30.85 and raises the RA symmetry score from 0.882 to 0.909, supporting the claim that functional constraints can be injected without retraining. The paper sees this as a step toward embedding physical and functional requirements into generative design workflows more broadly.

What carries the argument

The symmetrizer is the load-bearing object: a linear operator R that projects an image onto the set of n-fold rotationally symmetric images. Two instantiations are proposed: Rotate-and-Average (RA), which computes x* = (1/n) Σ_{k=1..n} A_k x averaging the image over n rotations, and Select-Sector-and-Stitch (SS), which selects one radial sector and replicates it to form the full image. At every DDIM denoising step t, the current latent z_t is decoded to x_t, symmetrized to R(x_t), re-encoded to z_t^r, and pooled with the original latent via z′_t = (1−λ) z_t + λ z_t^r, where λ = s(z_t, z_t^r) w / t^d, with s the normalized cosine similarity, w a weight, and d a decay constant. This projection-plus-decayed-pooling is what carries the argument: it injects the functional constraint while limiting the noise that full symmetrization would introduce.

What would settle it

Run an independent physical evaluation of the generated wheel designs—e.g., finite-element stress analysis or a rotational-balance test—and check whether designs produced by FIT (RA) actually show better stress distribution or rotational stability than the baseline interpolations. If the higher pixel symmetry scores do not correspond to better physical performance, the central claim that symmetry guidance improves functional compliance is falsified.

Watch

Extended reading notes

Core claim

The central claim is that enforcing rotational symmetry as a hard projection during the denoising steps of latent diffusion interpolation makes the generated wheels both more realistic, as measured by FID, and more compliant with the constraint, as measured by symmetry scores, compared with the unconstrained diffusion interpolation baseline of Wang and Golland (2023). The discovery is that a zero-shot, training-free constraint can be embedded directly into the inference loop by alternating decoding, symmetrization, re-encoding, and pooling, with the pooling strength modulated by cosine similarity between the original and regularized latents and decayed over steps. This is demonstrated on a dataset of 1,439 real wheel images with 5,700 interpolations, where FIT (RA) lowers validation FID from 39.28 to 30.85 and FIT (SS) from 39.28 to 33.98, and simultaneously increases the symmetry score from 0.882 to 0.909 (RA) or 0.507 to 0.601 (SS). The paper interprets this as evidence that physical functional requirements can be imposed during generation without additional training data or labels.

Load-bearing premise

The load-bearing assumption is that pixel-level rotational symmetry, as defined by the RA or SS symmetrizer, is a valid proxy for the physical functionality of a wheel, and that projecting the image with that same operator during denoising improves true functional validity.

Editorial extensions

If this is right

  • If the central claim holds, generative design tools can inject specific physical constraints into interpolation generation with no retraining, by plugging in a suitable differentiable regularizer.
  • The FID and symmetry gains reported suggest that functional constraints can serve as a regularizer for realism, not just a restriction, since the constrained outputs beat the unconstrained baseline on realism.
  • The method's zero-shot nature means it can be applied to any pretrained latent diffusion model and any constraint that can be expressed as a projection on image pixels.
  • The paper's decayed-pooling mechanism offers a general recipe for balancing constraint satisfaction against generation quality, which may be transferable to other guided-diffusion settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A caveat the paper itself notes is that the reported symmetry metric is computed from the same symmetrizer operator used for guidance, so the symmetry gains partly reflect the optimization target; an external, independent functional measure would be needed to confirm that pixel symmetry translates to physical functionality.
  • The cosine-similarity weighting and decay are manually chosen hyperparameters; one can infer that making these adaptive per step or per image might further improve the realism-compliance tradeoff, but the paper does not explore this.
  • The framework generalizes beyond wheels to any visual regularizer encodable as a projection into a constraint set, so the paper's case study is a template rather than a single-purpose fix.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces FIT (Functional Constraints in InTerpolation), a zero-shot inference-time method for generating interpolated wheel designs with enforced rotational symmetry. Given two wheel images, the method uses a latent diffusion model with DDIM sampling; at each denoising step it decodes the current latent, applies one of two symmetrizers (Rotate-and-Average RA or Select-Sector-and-Stitch SS), re-encodes the regularized image, and blends the two latents with weight λ = s(z,z^r) w/t^d based on cosine similarity and a decay schedule. Experiments compare with the Wang and Golland interpolation baseline on an EVOX wheel dataset, reporting lower FID for FIT(SS) and FIT(RA) (Table 1) and higher symmetry scores for the matching regularizer (Table 2). Ablations (Tables 3 and 4) show that similarity weighting and decay improve FID and that applying the constraint only at the end hurts RA.

Significance. The problem—enforcing a physical constraint during diffusion-based design interpolation without retraining—is timely, and the FID evidence is externally grounded: the generated wheels are compared with 1,439 real wheel images, and the reported gains (39.28→30.85 on validation; 46.01→33.38 on test) are large. The method is training-free, and the ablation in Table 3 shows that the decay and similarity mechanisms matter. However, the functional-compliance claim currently rests on a symmetry score that is defined with the same projection operator used during generation, so the paper does not yet establish that FIT improves true physical functionality; the RA formalization also needs correction. With an independent validation and a corrected definition, this would be a solid and useful contribution.

major comments (3)
  1. [5.3, Eq. (Sym), Table 2] The central claim that FIT 'generates designs that more closely satisfy physical and functional requirements' is not supported by an independent metric. The symmetry score Sym = 1 − ||R(x)−x||/K is computed with exactly the same operator R that FIT applies in Eqs. (2)–(4), so the improvement in Table 2 is at least partly a consequence of optimizing the evaluation metric itself. The pattern in Table 2 is consistent with this: FIT(RA) improves Sym(RA) (0.882→0.909) while Sym(SS) moves only 0.507→0.519, and FIT(SS) improves Sym(SS) (0.507→0.601) while Sym(RA) drops (0.882→0.867). Please add an external functional-compliance evaluation (e.g., human expert ratings, spoke-count or radial-geometry measures, or a physics-based surrogate) and report error bars or significance tests. Without this, the functional-compliance conclusion cannot be distinguished from an optimization artifact.
  2. [4.2] The RA regularizer as written does not formalize n-fold rotational symmetry. The text states that 'each A_k represents a rotation by 2π/k' and defines symmetry as A_i x = A_j x for all i,j ∈ [n]. The n-fold rotation group consists of rotations by 2πm/n for m=0,...,n−1, not rotations by 2π/k for k=1,...,n. For odd n, the stated condition requires invariance under a π rotation (k=2) and other angles that a genuine n-fold symmetric object need not have, and the averaging projection x* = (1/n)Σ_{k=1}^n A_k x does not generally project onto the subspace of n-fold symmetric images. Please correct the definition (rotations by 2πk/n for k=0,...,n−1) and state the value of n used in the experiments, or the reported RA results are not tied to the stated physical property.
  3. [5.1, 4.3] The quantitative results are not reproducible as reported because the hyperparameters and schedules are not given. The method depends on the symmetry weight w, the decay constant d, the fold count n, the number of DDIM steps, and the exact functional form of the decay; Section 5.1 only says these are tuned on the validation set. Please report the selected values for w, d, n, the DDIM step count, and the search range, and provide error bars (e.g., bootstrap over generated samples or multiple seeds) for the FID and symmetry numbers in Tables 1 and 2.
minor comments (6)
  1. [4 (opening)] The sentence 'The proposed system encodesFunctional Constraints inInTerpolation (FIT)' is missing spaces; the acronym should be introduced with proper spacing.
  2. [4.3] After Eq. (4), the phrase 'the the pooling' contains a duplicated article.
  3. [5.3] Reusing the notation R(x_t) for both the regularizer in Eqs. (2)–(4) and the operator inside the symmetry score obscures the circularity; please introduce a distinct notation for the evaluation operator.
  4. [Figure 2 caption] The caption contains 'a interpolated combination'; it should be 'an interpolated combination'.
  5. [Table 4 caption] The phrase 'applied with the constraints at the end of the interpolation' is ambiguous; it should be 'applied with the constraint at the end of the generation process'.
  6. [Limitations paragraph] The sentence 'Functionality and performance are not necessary reflected in visual patterns' should read 'not necessarily reflected in visual patterns'.

Circularity Check

1 steps flagged · score 6.0 of 10

Functional-compliance evidence is circular: the Sym metric is defined from the same symmetrizer R that FIT applies during denoising, so Table 2 measures the method's own projection rather than independent physical validity.

  1. self definitional [Section 4.3 (Eqs. 1-4) and Section 5.3 (Sym definition)]
    "zr_t = E(R(xt)) (2) ... z′_t = (1− λ)·zt + λ·zr_t (4) ... Then, our symmetry score is computed as Sym = 1− ||R(xt) − xt||/K where K is a normalization constant equal to the square root of the product of dimensions of xt."

    FIT's denoising step (Eq. 4) takes a convex combination of the latent zt and the latent of the symmetrized image E(R(xt)), so every step moves the generated image toward R(xt) and mechanically shrinks ||R(xt) − xt||. Section 5.3 then defines the success metric Sym as exactly 1 − ||R(xt) − xt||/K using the same operator R. Thus Table 2's reported Sym improvements are largely the method applying its own regularizer; they are not evidence from an independent check that the wheels are physically functional. The specificity of Table 2 (FIT(SS) mainly raises Sym(SS), FIT(RA) mainly raises Sym(RA)) confirms that each variant is optimizing its own evaluation metric.

full rationale

The paper has two families of claims. The realism claim (lower FID) is supported by comparison against real wheel images and is not circular. The functional-compliance claim, however, rests on the Sym scores in Table 2, and those scores are defined using the very symmetrizer R that FIT applies during generation (Eq. 4). Because z′_t is a convex combination of zt and E(R(xt)), the method directly minimizes the same distance ||R(xt)−xt|| that Section 5.3 uses as Sym. So Table 2 demonstrates that FIT enforces its own projection, not that an independent physical/functional test is satisfied. This is a classic self-definitional evaluation: the metric is the optimization target. The cross-metric pattern in Table 2 reinforces the point. No load-bearing self-citation or imported uniqueness theorem appears; the related-work citations to the authors' own prior papers are not used to justify the core derivation. A separate mathematical concern (not circularity) is that Section 4.2 defines RA using rotations by 2π/k for k=1..n, which literally is not the n-fold rotation group; as written it would not provably project onto n-fold symmetric images. Also, the paper's own Limitations paragraph concedes visual symmetry need not imply material/functional soundness. Overall, partial circularity: the central functional claim is largely self-referential, while the FID realism result remains independent evidence.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its contribution is algorithmic: two symmetrizer operators, a pooling rule, and a decay schedule. The main hand-chosen quantities are the regularization weight, the decay constant, and the number of rotational sectors, none of which are fully specified in the text.

free parameters (3)
  • w (symmetry weight) = 0.25, 0.35 (validation-tuned; swept 0.1-0.4 in Fig. 3)
    Controls the strength of the regularization in Eq. 3; selected on the validation set.
  • d (decay constant) = not specified in text
    Controls how quickly regularization impact decays over DDIM steps; tuned on validation, but the exact value is not reported.
  • n (fold symmetry) = not stated
    Number of rotational sectors used by RA and SS; essential to both symmetrizers, chosen by hand, and not reported in the paper.
assumptions (5)
  • standard math The symmetrizer projection maps any image into the subspace of n-fold rotationally symmetric images.
    RA averages the n rotated copies; SS replicates a chosen sector, so both are linear projections by construction (Section 4.2).
  • domain assumption Rotational symmetry is an essential functional requirement for automotive wheels.
    Stated in the abstract and Section 4; accepted for the case study but not derived from physics.
  • domain assumption Stable Diffusion 1.5, together with the latent interpolation schedule of Wang et al., produces meaningful wheel interpolations.
    The paper adopts this backbone and noise schedule without independently validating them on wheel images (Section 4.1, Section 5.1).
  • domain assumption Decoding, regularizing, and re-encoding the image at each step preserves enough information for useful guidance.
    The paper assumes the round trip xt = D(zt), z_t^r = E(R(xt)) does not destroy latent semantics; only FID is used to check this empirically (Eqs. 1-2).
  • ad hoc to paper The weight formula lambda = s(z_t,z_t^r) w / t^d is a suitable way to balance constraint and realism.
    Introduced in Section 4.3 with no derivation; w and d are tuned on the validation set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stylish and Functional: Guided Interpolation Subject to Physical Constraints." pith.science (2026). https://pith.science/paper/CNRBTVUT

@misc{pith2026241215507,
  author       = {Pith},
  title        = {Pith review of: Stylish and Functional: Guided Interpolation Subject to Physical Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CNRBTVUT}},
  note         = {Machine review of arXiv:2412.15507}
}
read the original abstract

Generative AI is revolutionizing engineering design practices by enabling rapid prototyping and manipulation of designs. One example of design manipulation involves taking two reference design images and using them as prompts to generate a design image that combines aspects of both. Real engineering designs have physical constraints and functional requirements in addition to aesthetic design considerations. Internet-scale foundation models commonly used for image generation, however, are unable to take these physical constraints and functional requirements into consideration as part of the generation process. We consider the problem of generating a design inspired by two input designs, and propose a zero-shot framework toward enforcing physical, functional requirements over the generation process by leveraging a pretrained diffusion model as the backbone. As a case study, we consider the example of rotational symmetry in generation of wheel designs. Automotive wheels are required to be rotationally symmetric for physical stability. We formulate the requirement of rotational symmetry by the use of a symmetrizer, and we use this symmetrizer to guide the diffusion process towards symmetric wheel generations. Our experimental results find that the proposed approach makes generated interpolations with higher realism than methods in related work, as evaluated by Fr\'echet inception distance (FID). We also find that our approach generates designs that more closely satisfy physical and functional requirements than generating without the symmetry guidance.

Figures

Figures reproduced from arXiv: 2412.15507 by the authors.

Figure 1
Figure 1. System overview: The proposed model takes a pair of reference images as the input [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Generating a interpolated combination of two different image sources can create distortion [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FID and symmetry score of the generated images over different weight of symmetry (w) in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Examples of images generated by the baseline [21] (top row) and FIT (bottom row). [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Example images generated by FIT (RA): Left: Applying the constraint during interpolation [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 26 canonical work pages

  1. [1]

    In International Conference on Learning Representations, 2023

    Zero-shot image restoration using denoising diffusion null-space model. In International Conference on Learning Representations, 2023

    work page 2023

  2. [2]

    Designaid: Using generative ai and semantic diversity for design inspiration

    Alice Cai, Steven R Rick, Jennifer L Heyman, Yanxia Zhang, Alexandre Filipowicz, Matthew Hong, Matt Klenk, and Thomas Malone. Designaid: Using generative ai and semantic diversity for design inspiration. In Proceedings of The ACM Collective Intelligence Conference , CI ’23, page 1–11. Association for Computing Machinery, 2023

    work page 2023

  3. [3]

    Ilvr: Conditioning method for denoising diffusion probabilistic models

    Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF Interna- tional Conference on Computer Vision (ICCV) , 2021

    work page 2021

  4. [4]

    Fashioning the future: Unlocking the creative potential of deep generative models for design space exploration

    Richard Lee Davis, Thiemo Wambsganss, Wei Jiang, Kevin Gonyop Kim, Tanja Käser, and Pierre Dillenbourg. Fashioning the future: Unlocking the creative potential of deep generative models for design space exploration. In Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems , CHI EA ’23. Association for Computing Machinery, 2023

    work page 2023

  5. [5]

    A study of the design process – the double diamond, 2005

    Design Council. A study of the design process – the double diamond, 2005

    work page 2005

  6. [6]

    Smooth diffusion: Crafting smooth latent spaces in diffusion models

    Jiayi Guo, Xingqian Xu, Yifan Pu, Zanlin Ni, Chaofei Wang, Manushree Vasu, Shiji Song, Gao Huang, and Humphrey Shi. Smooth diffusion: Crafting smooth latent spaces in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  7. [7]

    Consumer response to overstyling: Balancing aesthetics and functionality in product design

    Henrik Hagtvedt and Vanessa Patrick. Consumer response to overstyling: Balancing aesthetics and functionality in product design. Psychology and Marketing, 2014

    work page 2014

  8. [8]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. InProceedings of the 31st International Conference on Neural Information Processing Systems , NIPS’17, page 6629–6640, 2017

    work page 2017

Show all 28 references
  1. [9]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations , 2022

    2022

  2. [10]

    A survey on generative adversarial networks: Variants, applications, and training

    Abdul Jabbar, Xi Li, and Bourahla Omar. A survey on generative adversarial networks: Variants, applications, and training. ACM Comput. Surv., 2021

    2021

  3. [11]

    Hong, Yan-Ying Chen, Kalani Murakami, Jonathan Q

    Youngseung Jeon, Matthew K. Hong, Yan-Ying Chen, Kalani Murakami, Jonathan Q. Li, Xiang Anthony Chen, and Matthew Klenk. Weaving ml with human aesthetic assessments to augment design space exploration: An automotive wheel design case study. In Extended Abstracts of the 2024 CH...

    2024

  4. [12]

    A style-based generator architecture for genera- tive adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for genera- tive adversarial networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396–4405, 2019

    2019

  5. [13]

    Pixel-level image fusion: A survey of the state of the art

    Shutao Li, Xudong Kang, Leyuan Fang, Jianwen Hu, and Haitao Yin. Pixel-level image fusion: A survey of the state of the art. information Fusion, 33:100–112, 2017. 8

    2017

  6. [14]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning , volume 139, pages 8821–8831, 18–24 Jul 2021

    2021

  7. [15]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022

    2022

  8. [16]

    Stabilizing training of generative adversarial networks through regularization

    Kevin Roth, Aurelien Lucchi, Sebastian Nowozin, and Thomas Hofmann. Stabilizing training of generative adversarial networks through regularization. In Proceedings of the 31st International Conference on Neural Information Processing Systems , 2017

    2017

  9. [17]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023

    2023

  10. [18]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations , 2021

    2021

  11. [19]

    Medical image interpolation using recurrent type-2 fuzzy neural network

    Jafar Tavoosi, Chunwei Zhang, Ardashir Mohammadzadeh, Saleh Mobayen, and Amir Mosavi. Medical image interpolation using recurrent type-2 fuzzy neural network. Frontiers in Neuroin- formatics, 15, 2021

    2021

  12. [20]

    Vandal and Ramakrishna R

    Thomas J. Vandal and Ramakrishna R. Nemani. Temporal interpolation of geostationary satellite imagery with optical flow. IEEE Transactions on Neural Networks and Learning Systems , 2023

    2023

  13. [21]

    Wang and Polina Golland

    Clinton J. Wang and Polina Golland. Interpolating between images with diffusion models. In ICML 2023 Workshop on Challenges of Deploying Generative AI , 2023

    2023

  14. [22]

    Image morphing: a survey

    George Wolberg. Image morphing: a survey. The visual computer, 14(8-9):360–372, 1998

    1998

  15. [23]

    Gan inversion: A survey

    Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(3):3121–3138, 2023

    2023

  16. [24]

    Transferable and principled efficiency for open-vocabulary segmentation

    Jingxuan Xu, Wuyang Chen, Yao Zhao, and Yunchao Wei. Transferable and principled efficiency for open-vocabulary segmentation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  17. [25]

    Thomas Gale V oho Seo Hooman Shayani Ye Wang, Nicole B. Damen. Inspired by ai? a novel generative ai system to assist conceptual automotive design. In Proceedings of the ASME International Design Engineering Technical Conferences and Computers and Information in Engineering Co...

    2024

  18. [26]

    Diffmorpher: Unleashing the capability of diffusion models for image morphing

    Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xingang Pan. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912–7921, 2024

    2024

  19. [27]

    Deep image blending

    Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , March 2020

    2020

  20. [28]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 3813–3824, 2023. 9

    2023

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.