Inspect PRISM judgments at paper level.

In this work, the authors propose to learn a universal simulator (UniSim) of real-world interaction through generative modeling (a diffusion model for outputting the next frame given the previous frame and the input actions).

They achieve so by careful orchestration of diverse datasets, which are rich along completely different axes (e.g., some videos have object-level diversity, some have densely labeled language instructions, and some have scene-level diversity).

Reasonably scalable approach to collect training data for the proposed simulator

The use of diffusion models to fuse different aspects of the diverse datasets with decent results is impressive

This paper presents UniSim, a video prediction and generative model aiming for serving as a universal simulator of diverse scenarios conditioned on input language-described actions.

It devotes a big effort in combining dataset with different modalities and information axes, trained a unified generative model, and shows the trained model can be used for downstream policy learning.

Huge effort devoted in unifying multiple large scale datasets

This paper introduces a universal simulator (UniSim) that aims to simulate how humans and agents interact with the world.

The proposed framework combines various types of datasets, including internet text-image pairs and robotics data, with the motivation that existing datasets are useful along different axes.

The paper uses a video diffusion model as an interactive simulator of the world.

They show applications of the proposed simulator such as training long-horizon embodied planners and low-level object manipulators.

While this work shows great promise in a range of downstream applications, I believe it might need more experimental evidence to support the claim that it can simulate low-level actions well.

Specifically, section 4.2 only shows results for a relatively simple object (mostly blocks) re-arrangement (without grasping, e.g.) on a table.

It will give us insights as to how fine-grained the controls are supported by the proposed simulator, even if it cannot simulate low-level actions perfectly.

Experiments demonstrated effectiveness for downstream policy learning

UniSim can simulate both high-level instructions and low-level control, which show zero-shot transferability to real-world scenarios, addressing the sim-to-real transferability problem.

It would be nice if the paper delved more into the limitations of the models.

The paper has shown that exciting results can be obtained, but it's useful for the community to know the limits of the generalization capabilities, especially if people want to use this in the future for various applications.

For reproducibility, it would be helpful if the authors could release the code and some example pre-trained checkpoints.

Particularly the sim-to-real transfer is a promising direction for using the proposed real-world simulator.

Very cool and impressive research direction and proposed method

I think the paper presents a very important step towards learning a universal video predictive world model.

The authors highlight the potential for UniSim to be used in broader applications, such as video captioning and rare event detection.

This is an interesting paper that presents some exciting results.

The paper is well organized and well-written.

The paper introduces a novel approach to learning a universal simulator (UniSim) of real-world interaction through generative modeling, integrating various datasets including internet text-image pairs, robotics, human activities, panorama scans, and simulated data.

The paper presents a novel approach to learning a universal simulator (UniSim) of real-world interaction, integrating diverse datasets in a conditional video generation framework.

UniSim is formulated as an observation prediction model, approximating sampling in a POMDP, and is trained using a video diffusion model.

The methodology of UniSim is well-explained, with clear illustrations of the training and inference processes.

The paper does not discuss the limitations of the proposed method, which is crucial for understanding its applicability and potential drawbacks.

The model's capabilities are demonstrated through its application in training embodied planners, low-level control policies, and video captioning models, showing potential in sim-to-real transfer.

The application of UniSim is demonstrated across various domains, including embodied planners, low-level control policies, and video captioning models, showcasing its versatility.

The paper includes several examples of UniSim's application, such as training an embodied planner, a low-level control policy, and a video captioning model, demonstrating its effectiveness.

There is a lack of quantitative evaluation, which makes it difficult to assess the performance of UniSim objectively.

However, the paper faces criticism for its limited evaluation, lack of comprehensive comparisons, and unclear presentation, particularly in the methodology and experimental setup.

The paper is well-written, making it easy to follow, and includes a comprehensive literature review.

The methodology and experimental setup are not clearly presented, particularly the training details and the generation process of UniSim.

The presentation of the paper could be improved, particularly in sections where the methodology and experimental setup are described.

There is a need for more detailed explanations and examples, especially in the introduction and application sections, to enhance reader comprehension.

The paper lacks comprehensive comparisons with other existing methods for learning real-world simulators, which could help in understanding the novelty and effectiveness of UniSim.

2 fair

2 fair

2 fair

3 reject, not good enough

Decision: Reject

Reasons: The paper, while presenting an innovative approach to learning a universal simulator (UniSim) of real-world interaction, falls short in several critical areas. The primary concerns include limited evaluation, lack of comprehensive comparisons, and unclear presentation, particularly in the methodology and experimental setup. These issues make it difficult to assess the robustness and effectiveness of the proposed method. Furthermore, the paper does not adequately address the limitations of the method, whic...

The paper introduces UniSim, a universal simulator of real-world interaction, designed to generate realistic visual outcomes of both high-level instructions and low-level controls.

By combining diverse datasets—spanning text-image pairs, robotics, navigation, human activities, and simulations—UniSim is trained within a conditional video generation framework.

The simulator is also proposed as an observation prediction model approximating sampling in a POMDP, allowing for long-horizon interactions.

The paper presents a sound conceptual framework and demonstrates promising results in training embodied agents and simulating real-world interactions.

The paper demonstrates that UniSim can be used to train embodied vision-language planners, low-level reinforcement learning policies, and video captioning models, enabling zero-shot real-world deployment.

The work highlights the potential of UniSim to bridge the sim-to-real gap and enable applications such as rare event simulation and embodied learning.

The paper presents a novel and ambitious vision for a universal real-world simulator, addressing a significant challenge in generative modeling and embodied AI.

The integration of diverse data sources into a single framework is a notable technical contribution, and the demonstration of zero-shot real-world deployment of trained policies is promising.

The use of UniSim as an observation prediction model in a POMDP setting is an innovative approach to simulating long-horizon interactions.

The paper also highlights a range of potential applications, from embodied learning to content creation, which underscores its relevance to multiple domains.

The paper makes a valuable contribution by proposing a novel approach to building a universal real-world simulator that integrates diverse data sources.

It introduces an innovative application of POMDPs for simulating long-horizon interactions and demonstrates the potential of UniSim in training embodied agents.

The paper presents a promising and novel idea with significant potential, but it lacks sufficient comparative analysis, detailed methodology, and comprehensive evaluation to fully establish its contribution and robustness.

With improvements in these areas, the paper could be accepted for publication.

While the paper presents a strong conceptual framework and promising results, the lack of detailed methodology and comparative evaluation introduces uncertainty in the assessment of its overall contribution and technical soundness.

The paper is well-structured and clearly written, with a logical flow from introduction to methodology and applications.

The review is based on a thorough analysis of the paper and the provided Q&A pairs.

This paper introduces *UniSim*, a universal real-world simulator constructed via a diffusion model trained on heterogeneous datasets spanning text-image pairs, robotics data, human activities, and simulations.

The paper successfully fuses internet text/image data (e.g., LAION-400M), robotics data (Bridge Data, RT-1), human activities (Ego4D, EPIC-KITCHENS), and simulations (Habitat, Language Table) into a single simulator.

Section 2.1 details how datasets are normalized (e.g., T5 embeddings for text, discretized controls for robotics), and Table 5 summarizes the data mixture weights.

The paper provides a detailed derivation of the diffusion model architecture (Section 2.2), including the use of classifier-free guidance (Ho & Salimans 2022) and multi-frame conditioning.

Equations (1)-(2) formalize the denoising process, and Table 6 includes specifics like optimizer settings, attention resolutions, and noise schedules.

The core innovation is to unify these disparate data sources into a conditional video generation framework that predicts observations ($ o_t $) based on actions ($ a_t $) and historical context ($ h_{t-1} $), effectively approximating a Partially Observable Markov Decision Process (POMDP).

This is a significant engineering feat, particularly given the heterogeneity of modalities (text, video, low-level controls).

The authors demonstrate UniSim's utility in training embodied planners, RL policies, and vision-language models, claiming zero-shot transfer to real-world robots and improved performance on video captioning tasks.

The authors show that policies trained entirely in UniSim can execute long-horizon tasks on real robots (Figure 7) and improve video captioning performance (Table 4).

These results suggest potential for reducing real-world data dependency in AI training.

Section 4.1 demonstrates zero-shot transfer for embodied planners, and Section 4.3 reports CIDEr improvements (+27.63 vs. 21.91 on MSR-VTT) for vision-language models trained solely on UniSim-generated data.

Include explicit comparisons in experiments (e.g., "Does UniSim outperform Godiva on long-horizon planning?").

While the technical proposal is compelling and the experiments demonstrate feasibility, the lack of rigorous benchmarking, statistical validation, and ethical considerations prevents a stronger rating.

The training hyperparameters (Table 6) and model architecture (Appendix C) are well-described.

Section 5 ("Related Work") briefly cites these works but does not analyze how UniSim differs or improves upon them.

The paper warrants acceptance with the understanding that substantial refinements are needed for broader impact.

The paper proposes UniSim, a video diffusion model that is able to condition on past frames and actions to forecast future frames.

It combines multiple datasets from various domains, including robot manipulation, robot navigation, human activities, and panorama scans.

The proposed method is applied to training an image-goal conditioned VLM policy, a VLM policy with low-level control actions, and training a video captioning model.

The proposed method is applied to a wide range of downstream tasks.

The results show that UniSim is able to generate high-quality videos and improve the performance of downstream tasks.

My primary concern is the lack of experimental details, which makes it hard to evaluate the contribution.

The paper states that UniSim is trained on a large amount of data from various domains, but it is unclear how much data is used from each domain.

Moreover, the paper does not provide any details on the training procedure, such as the training time, the number of GPUs used, and the optimization algorithm.

This lack of information makes it difficult to reproduce the results and to assess the significance of the proposed method.

The paper also lacks a thorough comparison with existing methods.

The paper does not compare UniSim with these methods, which makes it difficult to assess the advantages and disadvantages of the proposed approach.

Furthermore, the paper does not provide a detailed analysis of the performance of UniSim on different tasks.

For example, in the context of video captioning, the paper only reports the CIDEr score on the ActivityNet Captions dataset.

However, there are other metrics that could be used to evaluate the quality of the generated captions, such as BLEU, METEOR, and ROUGE.

Moreover, the paper does not provide any qualitative examples of the generated captions, which makes it difficult to assess the strengths and weaknesses of the proposed method [5].

The paper is well-written and easy to follow.

The paper proposes to combine multiple datasets, which is an interesting idea.

For example, in the context of training VLM policies, there are several methods that use diffusion models to generate data for training [1, 2, 3, 4].

The paper proposes a method to learn a simulator of the real world through generative modeling.

The simulator is trained on a variety of datasets, including internet text-image pairs, motion and action rich data from navigation, manipulation, human activities, robotics, and data from simulations and renderings.

The method is trained on a variety of datasets, including internet text-image pairs, motion and action rich data from navigation, manipulation, human activities, robotics, and data from simulations and renderings.

The simulator is used to train embodied planners, low-level control policies, and video captioning models, which can generalize to the real world.

The simulator is also used to simulate long-horizon interactions and to generate training data for other machine intelligence tasks.

The simulator is used to train embodied planners, low-level control policies, and video captioning models, which can generalize to the real world.

The simulator is also used to simulate long-horizon interactions and to generate training data for other machine intelligence tasks.

The paper proposes a novel method for learning a simulator of the real world through generative modeling.

5: marginally below the acceptance threshold

This paper investigates a fundamental topic in consistency models (CMs), specifically the challenges of discretization errors and the resulting training stability issue.

Novelty. This paper's novelty is evident in several aspects.

First, it studies an important but less studied problem: consistency models in continuous time, together with the training stability and discretization error of consistency models.

Model architecture modifications are original since existing works are mostly inherited from Diffusion Models' design and focus on the training techniques and formulations, leaving the architectural design underexplored.

S3 - The unified perspective on previous diffusion and flow-matching parameterizations is thorough, complete, and well-grounded, offering novel insights that could benefit the community.

Consistency Models can be trained in discrete or continuous time, either from scratch using a dataset or distilled from pretrained teacher scores.

While continuous-time CMs eliminate the discretization errors present in their discrete-time counterparts, they suffer from training instability, a problem that is not yet well understood in the research community.

This work conducts a comprehensive study into continuous-time CMs, covering forward process parameterization, network architecture, and training techniques.

The authors first develop a simplified diffusion process formulation called TrigFlow, which unifies EDM and Flow Matching for the first time.

Building upon this foundation, they analyze the gradient flow of continuous-time CMs, identify the root cause of training instability, and mitigate this issue through modifications to time embeddings and adaptive group normalization.

Additional training techniques, such as adaptive weighting functions and annealing, further contribute to improved training stability and scalability.

The proposed TrigFlow, as a novel unification of EDM and Flow Matching, substantially simplifies the analysis presented later and the practical techniques.

I particularly appreciate the in-depth investigation into the training dynamics and gradient analysis of continuous-time CMs.

Additionally, the paper discusses efficient and stable implementation strategies for continuous-time CMs.

The authors propose improvements to the consistency models generative paradigm and named their new method sCM.

Specifically they -vastly- improve the FID for consistency models with the introduction of several new ideas to both stabilize and simplify continuous consistency models.

My understanding is the main claim of simplification for sCM comes from the simplification of EDM (Kerras et al.) normalizing design, resulting in $c_{in}=c_{skip}=1$ which in turns simplifies the continuous expression of consistency models.

Another simplification is the combination of both EDM and Flow Matching concepts into their method which they call TrigFlow.

Yet another simplification, not claimed as such by authors, is the use of vanilla L2 loss compared to Huber/LPIPS use in previous iterations of consistency models.

This last simplification has the additional benefit to be more probabilistically grounded.

There are 3 main proposed ideas to stabilize the training of consistency models: 1. Identity-time transformation as a replacement to the log-transformation from EDM 2. Fourier embedding of the time dimension are replaced by positional embeddings 3. AdaGN is modified to also normalize the conditioning inputs for scale and bias.

More ideas are also proposed in the training objective to stabilize training, namely: tangent normalization and tangent warmup.

It is my understanding that the adaptative weighing is the same as in EDM.

The paper presents a unified perspective on diffusion-based and flow-based generative models and introduces a comprehensive set of techniques aimed at improving the training stability and overall performance of continuous-time consistency models for large-scale image generation.

The techniques include: 1) enhancing time transformation and embeddings, 2) replacing the AdaGN layer with Adaptive Double Normalization, 3) normalizing the tangent function and applying tangent warm-up, 4) implementing an adaptive weighting function in the training objective, and 5) optimizing forward-mode differentiation.

S1 - The paper provides a comprehensive analysis and set of solutions addressing the numerical instability issues in continuous-time consistency models, significantly improving performance and enabling the model to achieve competitive results on selected benchmarks.

W1 - Several design choices appear arbitrary and lack supporting evidence.

This work proposed a set of improved training techniques to stabilize the training of continuous-time consistency models, including new consistency function formulations, new network architectures and new training objectives.

This work proposed a new diffusion formulation, called TrigFlow, that unifies EDM and Flowing Matching, and also simplifies the analysis of continuous-time consistency models.

It provided a thorough analysis of the training stability of continuous-time consistency models, from the perspective of network architecture, training objective and diffusion process parameterization.

Although I really like the improvements of continuous-time consistency models, which could fundamentally eliminate the discretization error in discrete-time consistency models, it comes with more time and memory costs related to JVP computation in the loss function.

To this end, this work introduces JVP of Flash Attention to reduce the costs, which is great.

Why do we need adaptive weighting?

CMs' theoretical foundation elucidates the importance of controlling the discretization error and eventually achieving consistency in continuous time.

The gradient analysis of continuous-time objective reveals the root cause of instability. To the best of my knowledge, this is the first paper to establish the gradient analysis for CMs.

S2 - Many of the enhancements are supported by detailed theoretical justification and experimental results.

A minor issue: In line 266, should it be $c_{\text{noise}}(t) = \frac{1}{4} \log(\sigma_d \tan t)$?

The resulting method, sCT/sCD, allows continuous-time CMs to be trained at an unprecedented scale, scaling up to 1.5B parameters on ImageNet 512x512.

These results significantly narrow the performance gap between CMs and state-of-the-art diffusion models to less than 10% in FID, while matching or even surpassing adversarial methods and discrete/continuous autoregressive models in both performance and efficiency.

Experiments. Proposed techniques allow for training continuous-time Consistency Models (sCMs) at an unprecedented scale.

Experiment results are impressive, matching/outperforming adversarial approaches, score distillation, and recent autoregressive models.

Gradient variances have been carefully controlled via adaptive weighting and normalization techniques.

Comprehensively studying the scaling behaviors of sCMs under continuous-time training.

Comparisons with improved score distillation baseline using many methods developed in this work confirm the mode coverage of CMs.

The paper also provides ample ablations to demonstrate the effects and the reasoning motivating these 3 proposed improvements.

The analysis is based on understanding the causes of training instabilities by decomposing the loss, validating each component experimentally and proposing changes to solve the root causes.

The experimental results are also outstanding resulting in very significant gains, essentially taking consistency models within 10% of the SOTA for diffusion models.

These techniques mitigate the numerical instability issues in continuous-time consistency models and enable the model to achieve highly competitive performance in class-conditioned image generation.

For example, in Section 4.1, the authors discuss the preference for Adaptive Double Normalization over AdaGN, but there is no experimental evidence supporting this choice.

Similarly, in Section 4.2, the authors propose training with linear warm-up w.r.t the model's time derivative, yet no evidence is provided to demonstrate this choice’s effectiveness.

Furthermore, Figure 5(b) suggests that incorporating adaptive weighting in a two-step setting may lead to worse performance, while in the one-step setting, it only yields marginal improvement.

W2 - In Sections 4.1 and 5.2, the paper discusses the training compute of sCM. However, including a comparison of compute efficiency with other models (e.g., ECT [1]) would be more insightful.

With these new training techniques, the proposed method called sCMs outperformed all previous consistency models in terms of one-step and two-step FIDs.

Experiments on CIFAR-10, ImageNet-64 and ImageNet-512 demonstrate the effectiveness of the proposed method and the scalability of continuous-time consistency models.

Still, there may be a considerable gap between the continuous-time and discrete-time consistency models.

I wonder if the paper can provide a more detailed comparison between sCMs and the previous discrete-time consistency models - ECMs, in terms of the training convergence and memory cost.

There is no explanation for the phenomenon that sCT performs better than sCD on CIFAR-10 and ImageNet-64, but sCTs performs worse on ImageNet-512.

Any intuition of why sCT suffers from increased variance at larger scales?

There are no ablation study results on “Adaptive Double Normalization” except for claiming it “removes its instability in CM training”.

In Figure 5b, it looks like “w/o adaptive weighting” achieves better two-step FIDs than “w/ adaptive weighting” and very similar one-step FIDs to “w/ adaptive weighting”.

In Figure 5c, do discrete-time CMs have a constant number of time steps $N$ or a timestep schedule up to the maximum number of steps $N$?

If it is the former one, it seems to be a bit unfair to discrete-time CMs because the scheduling of time steps is very important to them.

Does it make more sense to compare with the best-performing discrete-time CMs?

In Figure 7, does the paper apply TTUR proposed by DMD2 (Yin et al. 2024a)?

Thus, a comparison with VSD + TTUR is more convincing.

In Figure 7, sCDs condition the consistency network on the guidance scale $s$.

I wonder if VSD also condition the generator on the guidance scale, for a consistent evaluation setting?

This is a very strong paper in analysis, practical techniques, writing, and experiment results.

Soundness. Its technical claims are well backed up by both theoretical analysis and empirical results.

Given the potential impact of this paper, I strongly recommend acceptance with conference highlights.

I did not find any apparent weaknesses in the analysis or experiments (including both ablation studies and performance evaluation).

There are research questions worth further investigation, as discussed below.

The paper is very well grounded mathematically and experimentally.

Presentation. The logical flow of this paper is well structured and smooth.

The problem statement is clearly defined, and the explanation of why discretization errors matter for CMs and the motivation toward continuous-time formulation is crystal clear.

The gradient analysis into continuous-time CMs is thoughtfully motivated and carefully organized.

Even the appendix is well-written, offering useful insights into the proposed techniques.

It was a great pleasure to read through the manuscript!

The mathematics while greatly simplified are still pretty complex and the paper shines in its clarity to make the logical reasoning easy to follow.

S4 - The paper is well-structured and easy to follow.

This paper is very well-written and easy to read.

From the DMD2 paper, TTUR improves the performance of VSD.

The paper introduces simplified, stabilized, and scalable continuous-time consistency models (sCMs) for few-step generative modeling.

It proposes TrigFlow, a new formulation unifying EDM and flow matching, simplifying diffusion models and their associated probability flow ODE and consistency models.

The paper analyzes instability in consistency model training, proposes a complete recipe to mitigate it, including improved time-conditioning, adaptive group normalization, and a re-formulated training objective.

Simplification and unification of diffusion model formulations through TrigFlow.

Mitigation of training instability with improved techniques.

It introduces novel techniques to simplify, stabilize, and scale up the training of these models, achieving state-of-the-art or competitive results.

These improvements lead to better performance in consistency training and distillation, achieving comparable or better results compared to previous discrete-time formulations.

The models, referred to as sCMs, demonstrate success across various datasets and model sizes, scaling effectively with increased compute and narrowing the FID gap with state-of-the-art diffusion models.

Achieving state-of-the-art or competitive results across different datasets and model sizes.

Effective scaling to large models on high-resolution datasets.

Soundness result: 4 (excellent)

Rating result: 7 (accept, but needs minor improvements)

Decision: Accept

The need for minor improvements in these areas makes the paper suitable for acceptance with appropriate revisions.

Presentation result: 4 (excellent)

Contribution result: 4 (excellent)

Reasons: The paper presents significant contributions to the field of few-step generative modeling, specifically in the context of continuous-time consistency models.

The paper introduces significant advancements in consistency models (CMs) for generative modeling, focusing on improving their training stability, scalability, and performance.

The introduction of TrigFlow offers a unified and simplified formulation that bridges existing methods, which is a significant contribution to the field.

The paper makes a meaningful contribution to the field of generative modeling by introducing TrigFlow and addressing key challenges in the training of continuous-time CMs.

The authors propose TrigFlow, a novel formulation that unifies existing diffusion model and flow matching approaches, simplifying the training of continuous-time CMs.

They address key challenges in CM training, such as instability and discretization errors, by introducing improved time-conditioning, adaptive group normalization, and a re-formulated training objective.

The paper presents a novel and technically sound approach to improving the training and performance of continuous-time consistency models.

The authors address critical issues in CM training, such as instability and discretization errors, with practical solutions like adaptive group normalization and improved time-conditioning.

The paper presents a technically sound approach with a clear methodology and empirical validation.

The proposed techniques for improving the training of continuous-time CMs are well-motivated and supported by experimental results.

The paper presents a novel and technically sound approach with promising empirical results.

While the technical contributions are clear, the lack of detailed methodology and theoretical discussion introduces some uncertainty regarding the full impact and reproducibility of the work.

The paper demonstrates that their proposed sCMs achieve comparable or better sample quality than previous discrete-time CMs and VSD methods, using significantly less sampling compute.

The results are validated on multiple datasets, including ImageNet 512×512, with a model size reaching 1.5 billion parameters, the largest CMs trained to date.

The work also highlights the advantages of continuous-time CMs over discrete-time variants and compares sCMs with VSD in terms of sample diversity and guidance compatibility.

The empirical results are compelling, showing that sCMs achieve high sample quality with reduced computational cost, particularly in two-step generation.

The paper also provides a comparative analysis with VSD and discrete-time CMs, highlighting the practical benefits of their approach in terms of sample diversity and guidance compatibility.

The empirical results demonstrate the effectiveness of the proposed approach, and the comparative analysis with existing methods adds value.

Despite its technical contributions, the paper lacks sufficient theoretical and practical discussion of the broader implications of its findings.

The methodology section is not detailed enough to ensure reproducibility, with missing information on hyperparameters, training settings, and computational resources.

The conclusions are partially supported by the evidence but lack explicit logical connections to the results, and some claims are made without direct reference to the supporting data.

Additionally, the paper does not adequately address the limitations of its approach or provide a comprehensive discussion of how these limitations might affect the broader field of generative modeling.

The paper is generally well-structured and provides a clear overview of the problem and proposed solutions.

The assessment is based on a thorough analysis of the paper's content and the provided Q&A pairs.

The paper introduces **sCM** (Simple, Stable, Scalable Consistency Models), a novel approach aimed at enhancing the training and performance of consistency models (CMs) in generative modeling.

The authors tackle the instability of continuous-time CMs through a series of technical innovations, including **TrigFlow**, a unified formulation of diffusion processes that integrates elements of EDM, flow matching, and velocity prediction.

Additional contributions include architectural improvements (such as **adaptive double normalization**) and training objective refinements (like **adaptive weighting**, **tangent normalization**, and **warmup**).

**Comprehensive Reformulation of Diffusion Processes:** The introduction of **TrigFlow** provides a unified mathematical formulation that seamlessly incorporates EDM, flow matching, and velocity prediction.

**Practical Contributions for Stability and Performance:** The authors present several practical techniques to enhance training stability and performance, such as **tangent normalization**, **adaptive weighting**, and **tangent warmup**.

**Clear Methodological Differentiation Between sCT and sCD:** The paper distinguishes clearly between **consistency training (sCT)** and **distillation (sCD)**, offering insights into their respective strengths and limitations.

**No Exploration of Alternative Encoders/Decoders:** The paper notes that the current encoder/decoder may not be optimal for consistency models but does not investigate alternative designs or architectures.

The paper contributes meaningful advancements in the training and stabilization of continuous-time CMs, particularly through TrigFlow and training objective refinements.

The paper makes a solid contribution to the development of continuous-time consistency models, introducing TrigFlow and several practical training enhancements that yield impressive empirical results.

The methodology is validated on large-scale datasets like ImageNet 512×512, where a 1.5B-parameter sCM model achieves competitive FID scores with far fewer sampling steps compared to traditional diffusion models.

The paper also compares sCMs with alternatives such as VSD, EDM, and other consistency-based methods, asserting superiority in performance and scalability.

These contribute to making continuous-time CMs viable for large-scale applications, as demonstrated in Figures 4 and 5.

**Empirical Evaluation Across Multiple Scales:** The paper demonstrates the scalability of sCMs across various model sizes and datasets, including a **1.5B-parameter model** on ImageNet 512×512, which is currently among the largest CMs ever trained.

The empirical results show that sCMs match or exceed the performance of established methods like VSD and EDM with fewer sampling steps, as seen in Tables 1 and 2.

For instance, sCT excels at smaller scales, while sCD maintains consistent performance across all scales, as illustrated in Figure 6.

**Systematic Comparison With Baselines:** The authors systematically compare sCMs with competing methods such as VSD, ECT, and EDM, highlighting the benefits of their approach in terms of sample quality and training efficiency.

**Overstatement of Claims Without Statistical Support (High Severity):** Several claims are made without proper statistical backing.

**Insufficient Ablation Studies (Medium Severity):** The paper provides limited ablation studies on individual components of the proposed method.

For example, the impact of **TrigFlow** alone versus in conjunction with other techniques (e.g., tangent normalization or adaptive weighting) is not thoroughly analyzed.

This weakens the ability to isolate the true contributions of each innovation.

**Lack of Confidence Intervals and Significance Testing:** Many of the reported performance gains (e.g., narrowing the FID gap) are presented without statistical rigor, making it difficult to judge their validity.

While the paper presents a compelling technical framework and robust empirical results, the lack of statistical rigor, incomplete documentation of hyperparameters, and insufficient ablation studies weaken the overall soundness of the claims.

The paper is reasonably confident in its claims, but the absence of statistical testing, ablation studies, and hyperparameter transparency leaves room for doubt regarding the reliability of the results.

This unification simplifies the parameterization of diffusion models and enables a cleaner theoretical treatment of the training objective, as seen in Equations (15)-(18).

In **Equation (6)**, the expression for the tangent function $ \frac{df_{\theta}^{-}(x_t, t)}{dt} $ involves $\sigma_d$, $F_\theta$, and time-dependent terms.

**Absence of Formal Proof for Unit Variance Independence:** The claim that the unit variance design renders the training objective independent of $\alpha_t$ and $\sigma_t$ is asserted but not formally proven.

**Ambiguity in Hyperparameter Settings (Medium Severity):** Critical hyperparameters such as `c` (used in tangent normalization), `H` (number of warmup iterations), and `P_mean`, `P_std` (proposal distribution parameters) are inconsistently documented across the paper.

For instance, in **Table 6**, the FID of EDM2-XXL is reported as 1.73, yet in **Table 2**, the same model is cited with an FID of 1.81 — a discrepancy that undermines reproducibility unless clarified.

**Incomplete Documentation of Hyperparameters:** Critical hyperparameters such as `c`, `H`, `P_mean`, and `P_std` are inconsistently reported, hindering replication of the experiments.

**Limited Discussion on Generalization Beyond Images (Low Severity):** While the paper focuses on image generation, it acknowledges limitations in extending sCMs to video generation or fine-grained tasks.

**No Analysis of Generalization to Video Generation or Fine-Grained Tasks:** The paper acknowledges potential limitations in extending sCMs to video generation but provides no experimental evidence or analysis to substantiate these claims.

Despite notable shortcomings in reproducibility and statistical rigor, the paper presents a valuable contribution to the field of generative modeling. The proposed methodologies are theoretically grounded, empirically supported, and offer promising directions for future research. Minor revisions to address the identified issues would strengthen the submission.

With appropriate revisions, the paper would merit acceptance.

This paper presents a new formulation called TrigFlow, which unifies EDM and Flow Matching, simplifying the formulation of diffusion models, probability flow ODE, and CMs.

The paper also addresses the training instability of continuous-time CMs by improving time-conditioning and using adaptive group normalization.

Additionally, the paper reformulates the training objective for continuous-time CMs, incorporating adaptive weighting and normalization of key terms, and progressive annealing for stable and scalable training.

The proposed TrigFlow formulation simplifies the formulation of diffusion models, probability flow ODE, and CMs, making it easier to understand and implement.

The paper addresses the training instability of continuous-time CMs, which has been a long-standing issue in the field.

The paper provides a complete recipe for mitigating the training instability of continuous-time CMs, including improved time-conditioning, adaptive group normalization, and a reformulated training objective.

The authors demonstrate the effectiveness of their methods through experiments on various datasets and model sizes, showing improved performance and scalability compared to previous methods.

The proposed methods are shown to be effective through experiments on various datasets and model sizes, demonstrating improved performance and scalability compared to previous methods.

The paper proposes a new formulation of consistency models (CM) that unifies EDM and flow matching.

The authors also propose a set of techniques to stabilize the training of continuous-time CMs.

The paper proposes a new formulation of consistency models that unifies EDM and flow matching.

The authors also propose a set of techniques to stabilize the training of continuous-time CMs.

The paper proposes a new formulation of consistency models that unifies EDM and flow matching.

The authors also propose a set of techniques to stabilize the training of continuous-time CMs.

The paper proposes a new formulation of consistency models that unifies EDM and flow matching.

The authors also propose a set of techniques to stabilize the training of continuous-time CMs.

The proposed method is evaluated on CIFAR-10, ImageNet 64×64, and ImageNet 512×512.

The method is compared to other consistency models and diffusion models.

The proposed method is evaluated on CIFAR-10, ImageNet 64×64, and ImageNet 512×512.

The method is compared to other consistency models and diffusion models.

The proposed method is evaluated on CIFAR-10, ImageNet 64×64, and ImageNet 512×512.

The method is compared to other consistency models and diffusion models.

The proposed method is evaluated on CIFAR-10, ImageNet 64×64, and ImageNet 512×512.

The method is compared to other consistency models and diffusion models.

The paper introduces the Exploratory Diffusion Model (ExDM) to address the exploration bottleneck in Unsupervised Reinforcement Learning (URL).

Unlike prior methods that use simple policies, ExDM leverages the superior expressive power of diffusion models to accurately model the complex and heterogeneous state distributions collected during exploration.

A diffusion model is trained on the replay buffer's state distribution.

A novel score-based intrinsic reward is calculated from this model's loss (its inability to fit a state), which guides the agent to under-visited regions.

To ensure efficiency, a simple Gaussian behavior policy is trained to maximize this intrinsic reward and is used for fast data collection, avoiding slow diffusion sampling.

The pre-trained Gaussian policy can be fine-tuned on downstream tasks using standard RL algorithms (like DDPG).

It was impressed that the decoupled training scheme (fast Gaussian actor, slow diffusion reward-calculator) is a clever and practical solution to the primary obstacle of using generative models in online RL: slow sampling speed.

This paper introduce the Exploratory Diffusion Model, which leverages the expressive power of diffusion models to fit diverse replay-buffer distributions, thus providing accurate density estimates and a score-based intrinsic reward that drives exploration into under-visited regions.

This mechanism substantially broadens state coverage and yields robust pre-trained policies.

Beyond exploration, ExDM develops an efficient decoupled training scheme and a fine-tuning algorithm for adapting pre-trained diffusion components to downstream tasks under limited interaction, with theoretical guarantees of convergence and optimality.

The authors proposed an unsupervised RL algorithm called ExDM with diffusion action head.

This work proposes a novel approach called ExDM for unsupervised RL using a diffusion model.

During pre-training, ExDM trains a diffusion model to model the state and action distributions of interactions with the environment and derives an intrinsic reward to encourage exploration that is inversely proportional with approximate probability of state visitations of the diffusion model.

Using these intrinsic rewards, the approach trains a Gaussian policy that explores the environment.

During fine-tuning, the Gaussian policy can be trained with task-specific rewards, or the diffusion policy trained during pre-training can be fine-tuned.

To enable fine-tuning of the diffusion policy, a novel regularized training objective is being derived, similar to soft RL and using implicit Q-learning from the offline RL literature to avoid out-of-distribution actions being sampled.

One aspect that dampens my otherwise very positive impression of the significance of this work is the unclear benefits of the diffusion policy part of the ExDM algorithm.

The diffusion policy is arguably the biggest and most complex novel contribution of this work, but it appears to not contribute meaningfully to the performance of the approach (see Weakness 1.).

Without the diffusion policy, the ExDM algorithm could have also "just" been a diffusion model of the state distribution to derive a slightly novel intrinsic reward with a Gaussian policy.

The introduction of this work significantly leans into the motivation that typical URL approaches use policies that are not sufficiently expressive (often discrete or Gaussian policies) to properly explore the environment during pre-training.

Theorem 4.1 as theoretical contribution of this work further supports this narrative and the pre-training and fine-tuning of the diffusion policy component of ExDM takes over large parts of Section 4.

However, despite this motivation and more expressive diffusion policy, the environment interactions during pre-training are still done only using the Gaussian policy (as per line 8 in Algorithm 1).

All this makes me question what the diffusion policy of ExDM truly adds to the method.

It appears the benefits of ExDM are not from training a more expressive policy in the diffusion policy, but from the diffusion model of state distributions that appears to provide a more informative intrinsic reward to exhaustively explore the environment.

The author addressed that this is the first work to successfully integrate diffusion models into the unsupervised exploration phase of RL.

The concept of using the diffusion model's density estimation loss as the intrinsic reward is a significant contribution over prior reward mechanisms (like RND or ICM).

I think that the paper provided a novel, non-trivial algorithm for fine-tuning the diffusion policy itself, complete with a formal proof of optimality (Theorem 4.2).

This goes beyond just using the model as a static prior.

Potentially general mechanism: A diffusion-based exploratory prior could be a broadly applicable way to induce diverse skills or state coverage that helps downstream RL fine-tuning and transfer.

The motivation is somewhat weak.

The approach proposed in this work appears original and novel.

While diffusion policies are not new, and diffusion models have been used to express various data distributions, their application to URL is novel to the best of my knowledge.

Furthermore, the theoretical contributions in Theorem 4.1 justifying the need for more expressive policies for unsupervised pre-training, and in deriving a novel algorithm for online fine-tuning of the diffusion policy are valuable to the community.

Overall, the method's superior performance is not marginal.

Its experiments dramatically outperform all baselines in complex exploration tasks (e.g., Fig. 2, where baselines get stuck and ExDM covers the entire maze) and shows consistent SOTA results across all aggregate metrics in URLB (Fig. 3).

There is a limitation in terms of performance gap: The paper's own experiments (Fig. 3) show that fine-tuning the simple Gaussian policy actually achieves better final performance than the proposed new, complex diffusion policy fine-tuning algorithm (Algorithm 2).

The reason should be explained and analyzed intensively.

Compared with Fig. 3(a) and (b), the expert normalized scores of the proposed algorithm in Fig. 3(c) were small.

The authors stated that the performance degradation may be due to limited interaction timesteps during fine-tuning.

While their new fine-tuning method (Algorithm 2) is a novel contribution, it is not yet fully optimized and is outperformed by a simpler, standard approach such as DDPG.

Therefore, it is expected that the paper's primary strength lies in its pre-training exploration (which produces a superior Gaussian policy) rather than its diffusion policy fine-tuning performance.

Empirical gains across multiple settings: The figure indicates consistent improvements over strong unsupervised exploration baselines in URL, in cross-embodiment transfer, and when initializing diffusion policies.

The performance seems to be very strong compared to baselines

The new approach is shown to lead to more exhaustive exploration, as measured by state coverage, during pre-training, and the work shows that fine-tuning of the pre-trained Gaussian and diffusion policies lead to higher performance compared to alternative pre-training approaches.

Similarly, the empirical results indicate a small but consistent improvement of ExDM compared to the strongest URL baselines.

Assuming these results were generated under fair hyperparameter tuning (see question 3), they demonstrate that ExDM is a significant contribution to the field.

The empirical evaluation also appears to follow good practice, and provides further ablations and analyzes to shed more light on the learned components.

Furthermore, fine-tuning of the Gaussian policy of ExDM still leads to higher performance than fine-tuning the diffusion policy (see Figure 3 (a) vs (c)), a fact that is acknowledged by the authors in Section 5.4.

Appendix C.3 states that "hyperparameters of baselines are taken from their implementations".

I would expect comparable effort to be spent on tuning hyperparameters across all approaches to have confidence in the empirical results presented in this work, and this should be clarified.

The baselines visualized in Figure 2 appear to be mostly poor performing or middle of the pack when looking at Table 1.

None of the strongest baselines (MEPOL, RE3, CIC) are included in Figure 2, supposedly to make the result of ExDM appear more impressive.

I would appreciate Figure 2 would show the strongest 1-2 baselines in each family which appear to be R3 and MEPOL for exploration and CiC for skill discovery baselines.