- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
E-Flow combines the style-content representation transport framework with the energy-gradient velocity parameterization and Langevin-based contrastive divergence, and introduces Meyer wavelet feature fusion. It has some improvements in metrics for representation disentanglement in style transfer.
Critical strengths
This work effectively integrates the style-content feature transfer framework of SCFlow with the energy modeling approach of Energy Matching. Additionally, it introduces energy modeling conditioned on temporal and content-style pairings, along with a feature fusion mechanism. The main contribution of this work is that E-Flow constructs a comprehensive style transfer framework. Experimental results demonstrate performance improvements on the SCFlow benchmark, showing the value of incorporating energy modeling into style transfer tasks.
Critical weaknesses
- The key components of E-Flow are mainly derived from existing work. Specifically, the style-content representation transport framework originates from SCFlow, while the energy-gradient velocity parameterization and Langevin-based contrastive divergence training have been explored in Energy Matching. Therefore, the main contribution of E-Flow is integrating these two existing approaches and applying them to the task of style transfer. Building on those, E-Flow further introduces some task-specific adaptations, including pair- and time-conditioned energy, local negative sampling around stylization targets, and Meyer wavelet-based feature fusion tailored for style transfer.
- The paper does not have a discussion and comparative experiments with CAtFM. As a follow-up to SCFlow, CAtFM incorporates contrastive supervision to improve style-content disentanglement. E-Flow requires a comparative analysis with CAtFM to show if the use of energy-gradient parameterization and Langevin CD offer additional advantages over simply adding a contrastive loss to SCFlow?
- As the main contribution, the effectiveness of the Meyer wavelet design has not been justified properly. Although low and high freq wavelet has interpretations in the spatial image domain as the semantic structure and texture, whether that can be extended to the coordinate dimensions of CLIP still needs to be justified. Also, in Table 2, some of the compared wavelets have better performance than the Meyer wavelet. Why Meyer was chosen as the final implementation has not been explained.
- The paper does not explicitly describe the backbone network architecture, model parameter count, training initialization, or optimization settings.
Questions for authors
- While the motivations for the stylization task adaptations mentioned are discussed, their methodological novelty relative to prior work could be more clearly established. It would be helpful if the authors explicitly distinguish the proposed technical contributions from the adaptation of existing techniques.
- The comparison experiments with CAtFM should be added.
- The questions related to the Meyer wavelet need to be further explained.
- Authors are recommended to specify details of the network architecture and training details. Overall, this work presents a novel framework for image style transfer. Although its innovation may not be groundbreaking, the proposed combination of existing work is intriguing and demonstrates performance improvements. I would be willing to raise my score if the authors could: 1) provide necessary comparisons and analysis for CAtFM; and 2) further explain the design choice of the Meyer wavelet.
3 comments
I appreciate the authors' detailed response and the additional experiments. However, I believe the core issues remain insufficiently addressed:
- While the authors have elaborated on the motivation for combining SCFlow with Energy Matching, their explanation primarily focuses on how existing techniques can be adapted for style transfer. The methodology contributions are still unclear.
- Regarding the comparison with CAtFM, these experiments do not answer my original question: what specific advantages do energy-gradient parameterization and Langevin CD offer over simply incorporating conventional contrastive supervision into SCFlow?
- Although the authors describe the wavelet fusion process, they have not addressed my primary concern: why should the low- and high-frequency components, which are obtained by applying wavelet decomposition to CLIP embedding coordinates, correspond to image structure and texture, respectively? While this logic holds in the spatial image domain, it cannot be directly generalized to the CLIP embedding space. The authors have restated this assumption without providing any supporting theoretical or empirical evidence. Overall, I remain unconvinced that the response sufficiently resolves these issues, and I therefore maintain my original rating.
Our previous responses focused too much on the details, which led to misunderstandings. We are very grateful for the opportunity to explain again.
Q1: What are the specific methodological contributions?
Unlike commonly used diffusion methods, we address style transfer from the perspective of an energy field. In recent years, movement along energy gradients has shown promise in generative modeling and contrastive learning. However, its direct application faces challenges involving non-invertibility, explicit modeling of the target distribution, and the absence of time conditioning.
Existing diffusion methods are also unsuitable, as they cannot simultaneously preserve content consistency and style controllability within an invertible framework. We therefore adopt a more suitable flow-based model.
Directly applying SCFlow leads to structural information loss because its concatenation-and-averaging operation acts as isotropic, non-invertible pooling. Once fused, the content and style codes cannot be cleanly separated and interfere with each other during reconstruction.
To address these issues, we
(1) Introduce a pair- and time-conditioned energy function, addressing the absence of time conditioning and dynamic modeling in the energy field. This enables the model to prioritize global structure early in generation and refine local texture at later stages.
(2) Introduce local negative sampling around the stylized target, addressing the overly global nature of style constraints, which can be diluted by the background or content objects. This focuses the contrastive signal on whether local regions appropriately express the target style.
(3) Introduce an invertible fusion interface based on Meyer wavelets, addressing the non-invertibility and isotropic nature of the fusion operation.
Together, these components form a style-transfer framework driven by energy, supported by flow as an invertible transport mechanism, and trained through local contrastive supervision. Its central feature is to shift the perspective on style transfer from denoising or statistical matching to descent along the energy gradient on an invertible manifold, thereby combining the generation quality of diffusion with the invertibility and controllability of flow.
Q2: What advantages do energy-gradient parameterization and Langevin CD offer over conventional contrastive supervision?
We apologize for not understanding your point in our previous response. This was our oversight. In fact, we conducted an experiment without Langevin CD, with the results shown below.
| Model | S10 ↑ | S25 ↑ | S51 ↑ | C25 ↑ | C200 ↑ |
|---|---|---|---|---|---|
| E-Flow with conventional supervision | .71/.35 | .67/.39 | .58/.38 | .74/.53 | .65/.56 |
| E-Flow with Langevin | .88/3.47 | .86/4.06 | .83/4.06 | .88/2.47 | .84/2.61 |
Compared with conventional contrastive supervision, our method offers the following advantages:
(1) Conventional contrastive supervision imposes static geometric constraints. Langevin CD, through energy gradients, acts directly on every intermediate state along the transport trajectory, embedding the contrastive signal into the velocity field itself and dynamically steering the model away from negative-sample regions during generation.
(2) Conventional contrastive supervision tends to learn global discriminative features and can overlook local patterns or multimodal structure in the data distribution. Langevin CD uses the stochastic exploration of Langevin dynamics to explicitly capture the local energy landscape, enforce the uniqueness of flows under different conditions, and thereby reduce overlap between conditions while better preserving local details.
(3) Conventional contrastive supervision often requires large numbers of negative samples to effectively constrain the distribution in high-dimensional generative tasks. Energy gradients provide more precise supervision, enabling stronger constraints with fewer negative samples. Combined with flow matching, this also enables high-quality sampling with fewer energy evaluations.
Q3: Why should low- and high-frequency wavelet components of CLIP embeddings correspond to image structure and texture?
Upon re-examining our explanation, we recognize that it contained a wording error. The low- and high-frequency components obtained by applying wavelet decomposition to CLIP embedding coordinates do not fully correspond to image structure and texture.
Our use of Meyer fusion actually addresses alignment at another level.
Through ODE integration, we achieve alignment in the time domain.
High-frequency alignment and low-frequency alignment provide another level of alignment, which is the intended purpose of introducing the wavelet transform.
Q1. Novelty in energy-guided style transfer
A1: We thank the reviewer. We build on two prior works: SCFlow for transport in CLIP space, energy matching for a field defined as the gradient of a scalar potential. But parts do not make an assembly; they must combine as they stand, and these do not. The combination breaks in four places, each from the task, not either method.
First, the object of the criterion. A potential over a data distribution asks whether a sample resembles it; style transfer asks whether this pair is appropriate at this state. With one target per pair there is no distribution, so a distributional objective degenerates; it must become a ranking, one positive against one negative refined by Langevin steps, since no invalid neighbour exists in the dataset and one must be made locally.
Second, conditioning. An unconditional potential scores a state identically under every pair, yet the same result is valid with one style reference and invalid with another. It must take the pair as an argument; then the field is autonomous on the joint state space, so only then are training contrast and inference integration one object.
Third, we connect the transport direction to a learned compatibility score. SCFlow directly predicts velocity, whereas E-Flow defines it as vθ(x,t)=−∇xEθ(x,t). Flow matching trains this gradient to move toward the paired target, while local contrastive divergence encourages lower energy for the target than for nearby negative states. The same scalar function therefore learns both how to move and which nearby states to favor.
Fourth, the endpoint. The criterion scores a joint state, but only decoder-accepted states can be rendered; inexact fusion makes the two diverge and drives the model somewhere undecodable. Mean fusion is that case; hence a complementary, invertible interface. Tab. 2 gives 3.12 against 3.87 to 4.16.
The four are interdependent: a criterion needs a valid endpoint, an endpoint an exact interface, an interface complementary targets, and targets a criterion that separates neighbours. Defaults anywhere discard the other three benefits. The drops are not marginal: without the energy parameterization, style FDR falls from 3.47–4.06 to 0.93–1.11; with mean fusion, to 3.12.
Q2. Comparative experiments and performance analysis
A2: Thank you for the reminder. Since CAtFM was working concurrently, this part was not included initially. We will add this content to the revised version. The comparison result is as follows.
Q3: Wavelet design and Meyer selection
A3: We thank the reviewer. On semantics, Figure 3 may cause confusion. In the revised version, we will make improvements. Fusion is carried out band by band: low adds to low and high adds to high, so the result in each band is determined jointly by the two sides at the same scale. The low band composes the content layout with the overall palette of the style, and the high band composes content edges with the style's texture and brushwork.
This also explains why the criterion has to be a scalar on the joint state. A one-sided score can tell whether a style resembles some style; it cannot tell whether the two bands compose.
Meyer is band-limited and smooth, so leakage between channels is small and the two supervision signals overlap least, which Haar cannot match. We do not claim it is best on the metrics — the four bases in Tab. 2 are close and all clearly beat mean — but that this family, orthogonal, frequency-selective and parameter-free, is the right fit.
Table 2 compares Mean fusion with dmey. Cells show NMI/FDR; S/C denote style/content class counts.
Q4: Architecture and training details
A4: Architecture/size. The Transformer has width 2048, 12 layers and 32 heads, operating on two 768-dimensional branches. Energy hidden/time dimensions are 2048/128. There are 633,968,129 trainable transport parameters: 618,999,296 in the backbone and 14,968,833 in the energy head. Frozen CLIP/unCLIP and duplicate EMA state are excluded.
Initialization/optimization. Trainable components are randomly initialized; pretrained CLIP/unCLIP weights remain fixed. Training uses Adam, learning rate 10−5, weight decay 0, batch size 384 and 16-bit mixed precision. We evaluate the model after 80,000 updates, using its online parameters. CD weight is .1; negative samples use Gaussian noise with σ=.05 and 20 Langevin steps, with step size decreasingfrom .01 to .002.