Pith. sign in

REVIEW 3 major objections 4 minor 8 references

Optimal transport can realign text and video embeddings so that annotation-free rewards for text-to-video post-training outperform human-preference-based methods.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:35 UTC pith:ATZ7K7CM

load-bearing objection Well-executed annotation-free T2V post-training with a genuinely new dual-OT reward module; the quality reward's text-conditioning is asserted rather than proven, but the paper is honest, thorough, and deserves a serious referee. the 3 major comments →

arxiv 2602.01624 v2 pith:ATZ7K7CM submitted 2026-02-02 cs.CV

PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards

Minh-Quan Le , Gaurav Mittal , Cheng Zhao , David Gu , Dimitris Samaras , Mei Chen This is my paper
classification cs.CV
keywords text-to-video generationreward-based post-trainingoptimal transportembedding alignmentsemantic groundingvideo-text matchingannotation-free alignmentquality reward
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to prove that reward-based post-training for text-to-video generation can work without any human preference annotations, provided the reward is evaluated in an embedding space realigned with optimal transport. It proposes PISCES, with two rewards: a distributional quality reward that transports text embeddings onto the real-video manifold before taking cosine similarity with generated videos, and a token-level semantic reward that uses a partial optimal-transport plan with spatio-temporal costs to sharpen cross-attention in a video-text matching classifier. The paper reports that this dual scheme outperforms both annotation-based and annotation-free post-training methods on short and long video generators, on automatic metrics and in a human preference study. If right, this removes the main scalability bottleneck of reward-based video post-training: collecting large preference datasets.

Core claim

The central claim is that the mismatch between pre-trained text embeddings and real-video embeddings, rather than the scarcity of paired preference data, is the bottleneck for annotation-free text-to-video post-training. PISCES corrects this mismatch in two complementary ways. At the distributional level, a neural optimal-transport map is trained to push text embeddings onto the real-video embedding distribution while preserving internal structure; the quality reward is then cosine similarity between the transported text summary embedding and the generated video's summary embedding, making the text embedding a proxy for a real video of that prompt. At the token level, a partial entropic opti

What carries the argument

The central object is the dual optimal-transport alignment: a distributional transport map between text and video embedding distributions, trained with squared-Euclidean cost through a two-network minimax objective; and a discrete partial optimal-transport plan over text tokens versus video patch tokens, with a cost that penalizes semantic dissimilarity, temporal distance, and spatial distance, computed with an entropic solver and fused into cross-attention in log space. The distributional map carries the quality reward by making a transported text embedding comparable to a generated video embedding on the real-video manifold; the token-level plan carries the semantic reward by selecting whi

Load-bearing premise

The quality reward assumes the transport map sends a given prompt's embedding close to a real video of that same prompt, but the map is trained only on the overall distributions of prompts and videos separately, with no pairing that ties a specific prompt to the video that matches it.

What would settle it

For prompts paired with both matching and non-matching videos, check whether the OT-transformed prompt embedding ranks the matching video's summary embedding closer than the non-matching one. If it does not, the quality reward is not text-conditioned, and gains attributed to the OT alignment would need to be explained by the semantic reward or by the consistency-loss regularizer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, annotation-free post-training can match or surpass human-preference-based post-training, removing the main scaling bottleneck for video alignment.
  • Quality and semantic supervision separate cleanly: the quality reward drives global realism and temporal coherence, the semantic reward drives object and action grounding, and combining them gives the best of both.
  • Because the same reward module works under direct backpropagation and reinforcement-learning fine-tuning, existing post-training stacks can adopt it without changing the optimizer.
  • The paper's cost analysis implies the optimal-transport map is a small, one-time addition relative to model pre-training, and the token-level transport plan doubles as an inspectable alignment of words to video regions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Not in the paper: if the transport map preserves only the aggregate distribution rather than each prompt's correspondence to its own real video, the quality reward could degenerate into a generic realism score; a per-item retrieval test would settle whether the OT alignment is doing text-conditioned work.
  • An editorial extension: the dual alignment recipe could transfer to other cross-modal post-training settings—image-text, audio-video, or embodied language—wherever frozen embeddings occupy misaligned marginal spaces.
  • The paper's Appendix K concedes that fine-grained token grounding is bounded by the frozen video-text encoder's spatial precision; reading forward, switching to a stronger base encoder should transfer directly into better grounding under PISCES.
  • A further implication: the log-space fusion of the transport plan with cross-attention is a structural prior that could be reused at inference time, for example to visualize which prompt words ground which video regions.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. PISCES proposes an annotation-free post-training method for text-to-video diffusion models, built around a Dual Optimal Transport-aligned Rewards module. A distributional OT map is trained from text and video embedding marginals and used in a cosine-similarity Quality Reward (Eq. 2), while a token-level partial-OT plan is injected into the cross-attention of an InternVideo2 VTM classifier to produce a Semantic Reward (Eq. 4). The paper reports VBench and human-preference gains over annotation-based and annotation-free baselines on VideoCrafter2 and HunyuanVideo, with direct-backpropagation and GRPO optimization variants, plus extensive appendices on ablations, hyperparameters, OOD behavior, and training cost.

Significance. If the central attribution is correct, this is a meaningful contribution: it suggests that carefully aligned embeddings can make annotation-free reward supervision competitive with or superior to human-preference-based post-training, which would improve scalability. The manuscript is unusually thorough in its appendices: Algorithms 1-3 are given in detail, the method is tested under two optimization paradigms, with two video backbones, two embedding encoders, a ViCLIP-based generalization check, an OOD prompt study, inter-rater reliability, and a discussion of reward hacking. These are strengths. However, the headline claim rests on a proxy in Eq. (2) that is not directly validated, on point estimates without uncertainty quantification, and on hyperparameters selected with the same VTM metric used as reward evidence. These load-bearing issues need to be resolved before the contribution is established.

major comments (3)
  1. [Section 3.1, Eq. (2)] The quality reward rests on the proxy T*(y[CLS]) ≈ x_real[CLS] for the same prompt. The OT map is trained on marginal distributions of text and video embeddings (Eq. (1), Algorithm 2) with no paired constraint tying a given prompt to its matching real video. OT is non-unique, and the checks in Table 4 and Fig. 6 (Mutual KNN, Spearman, t-SNE) are aggregate geometry measures; they do not establish that the specific target used in Eq. (2) is the paired or semantically equivalent real-video embedding. If T* only projects text onto the video manifold while ignoring prompt semantics, R_OT-quality is an unconditional realism reward, and the quality gains in Table 3 would be explained by target manifold placement rather than OT alignment. Add a control replacing T*(y[CLS]) with a prompt-independent real-video embedding (mean or random) and report a retrieval-based per-prompt correspondence metri
  2. [Section 4.3, Appendices F and J] The hyperparameters m=0.9, γ=η=0.2 are selected by maximizing Video-Text Matching (VTM) accuracy on WebVid10M (Appendix J), and the same VTM accuracy is then reported as evidence that partial OT and spatio-temporal constraints improve alignment (Table 7). Because the semantic reward is defined as the VTM classifier logit, this is hyperparameter selection on the evaluation metric, not independent validation. Report the chosen configurations' VBench Quality/Semantic scores on held-out prompts, or use an alignment metric separate from the reward design, to support generalization of these constraints.
  3. [Section 4.2, Tables 1-2 and Fig. 3] The headline claim that PISCES 'significantly outperforms' annotation-based methods is based on point estimates without error bars, repeated seeds, or significance tests. Some margins are small (e.g., VideoCrafter2 Semantic: 77.54 vs VideoDPO 77.38; Total: 82.75 vs 81.93 in Table 1). The human-preference percentages in Fig. 3 likewise lack confidence intervals, although inter-rater reliability is reported in Appendix L. Please report variance over at least three seeds or bootstrap confidence intervals, and appropriate significance tests for the main comparisons.
minor comments (4)
  1. [Section 3.2 / Appendix C] The mapping from the 'effective transported fraction m' to the unbalanced Sinkhorn relaxation parameters τ_a, τ_b is stated but not derived. A short explanation of why τ(ρ)=ρ/(ρ+ε) implements the desired mass fraction would help reproducibility.
  2. [Table 4] The Contrastive row leaves Spearman correlation empty. If not measured, say so explicitly; if it is undefined, explain why.
  3. [Appendix J] Figure 8 labels the x-axis 'Configuration Index' but the text does not provide a mapping from index values to (γ, η) pairs. Including the actual grid values on the axis would make the sensitivity result directly interpretable.
  4. [General] The term 'annotation-free' is used to mean 'no human preference annotations,' but the method still relies on paired text-video data (WebVid10M, VidGen-1M) and a pre-trained VTM classifier. Clarify this terminology early to avoid overclaiming.

Circularity Check

1 steps flagged

Semantic-reward validation is self-referential via the VTM classifier, but the central VBench/human comparisons are external.

specific steps
  1. fitted input called prediction [Eq. (4); Appendix F (Table 7); Appendix J]
    "ROT-semantic = softmax(VTM[÷x])_(idx=1) ... 'we measure the Video-Text Matching (VTM) accuracy on 10,000 WebVid10M ... video-text pairs using our OT-aligned Semantic Reward with InternVideo2' ... 'using Partial OT with a mass parameter m=0.9 achieves the best score of 89.36%, improving by +8.11% over vanilla cross-attention.'"

    The semantic reward is literally the VTM classifier's positive logit after POT-refined attention (Eq. 4). The hyperparameters (m, gamma, eta) used in that reward are selected by maximizing the same VTM accuracy on WebVid10M (Appendix J), and Appendix F/Table 7 then reports the VTM accuracy gain (+8.11%) as evidence that the POT component improves alignment. Since post-training directly optimizes this logit and the validation metric is the same classifier, the reported VTM-based validation is forced by the reward's definition rather than being an independent check. The central VBench/human comparisons are external, so the overall claim is not reduced, but this component's evidence is partially circular.

full rationale

The core PISCES claims—outperforming annotation-based and annotation-free methods on VBench and human preference—are benchmarked externally (VBench detectors, human raters), so the overall derivation is not equivalent to its inputs. The OT map training (Algorithm 2) is an unpaired distributional fit; the quality-reward assumption T*(y[CLS]) ≈ x_real[CLS] is an unverified approximation rather than a construction, and the absence of a prompt-independent control is a correctness/attribution gap, not circularity. The one concrete circular-by-construction element is the semantic reward's validation: Eq. (4) defines the reward as the VTM positive logit, and Appendix J selects the reward's hyperparameters by maximizing VTM accuracy on WebVid10M, while Appendix F reports that same VTM accuracy as evidence of improved alignment. That component-level evidence reduces to the reward's own definition. No load-bearing self-citation or imported uniqueness theorem appears; self-citations (e.g., Hummingbird) are background. Thus score 4: partial circularity in a component, central claim intact.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on four domain assumptions about the reliability of frozen VLM embeddings, the semantic fidelity of an unpaired OT map, the trustworthiness of the VTM classifier, and dataset transfer. The OT map's per-prompt semantic identity is the most fragile assumption. Free parameters m, γ, η are tuned on the same VTM accuracy used by the semantic reward, adding a mild fitting component. No new physical or mathematical entities are invented.

free parameters (4)
  • partial OT mass fraction m = 0.9
    Selected by sweeping VTM accuracy on WebVid10M (Appendix F); directly controls how many text tokens are allowed to be dropped and is the same objective the semantic reward maximizes.
  • temporal penalty weight γ = 0.2
    Tuned over [0.0,0.5] in Appendix J to maximize VTM accuracy; balances temporal mismatch cost in the token-level OT plan.
  • spatial penalty weight η = 0.2
    Tuned with the same Appendix J sweep; balances 2D-position mismatch cost in the token-level OT plan.
  • entropic temperature ε (Sinkhorn) = 0.05
    Set manually for the entropic/unbalanced Sinkhorn solver; stated in §C and Appendix C without a dedicated sensitivity analysis.
axioms (4)
  • domain assumption Pre-trained VLM embeddings (InternVideo2/ViCLIP) provide a valid representation space in which OT-aligned cosine similarity and VTM logits correlate with human judgment of video quality and text fidelity.
    The entire annotation-free strategy rests on this; used throughout §3 and §4.
  • ad hoc to paper A Neural OT map trained on text and video marginal distributions maps each text embedding to the corresponding real-video embedding while preserving semantic structure.
    Eq. (1)-(2) and Algorithm 2 train T on marginals without pairwise text-video supervision; per-item semantic correspondence is not guaranteed by OT. The paper only checks aggregate Spearman correlation (Table 4).
  • domain assumption InternVideo2's Video-Text Matching head is a reliable oracle for semantic grounding; optimizing its positive logit is a valid proxy for human semantic fidelity.
    Eq. (4) defines the semantic reward as the VTM positive logit. Appendix O acknowledges reward-hacking risk, and CD loss is used as a mitigation rather than an independent proof of alignment with human judgment.
  • domain assumption Optimizing on WebVid10M/VidGen-1M post-training data transfers to VBench and the 400 human-evaluation prompts.
    Stated in §4.1 and asserted in the generalization discussion; no cross-dataset training/evaluation leakage analysis is provided beyond OOD cosine checks in Appendix I.

pith-pipeline@v1.3.0-alltime-deepseek · 25166 in / 15120 out tokens · 158226 ms · 2026-08-03T05:35:41.622535+00:00 · methodology

0 comments
read the original abstract

Text-to-video (T2V) generation aims to synthesize videos with high visual quality and temporal consistency that are semantically aligned with input text. Reward-based post-training has emerged as a promising direction to improve the quality and semantic alignment of generated videos. However, recent methods either rely on large-scale human preference annotations or operate on misaligned embeddings from pre-trained vision-language models, leading to limited scalability or suboptimal supervision. We present $\texttt{PISCES}$, an annotation-free post-training algorithm that addresses these limitations via a novel Dual Optimal Transport (OT)-aligned Rewards module. To align reward signals with human judgment, $\texttt{PISCES}$ uses OT to bridge text and video embeddings at both distributional and discrete token levels, enabling reward supervision to fulfill two objectives: (i) a Distributional OT-aligned Quality Reward that captures overall visual quality and temporal coherence; and (ii) a Discrete Token-level OT-aligned Semantic Reward that enforces semantic, spatio-temporal correspondence between text and video tokens. To our knowledge, $\texttt{PISCES}$ is the first to improve annotation-free reward supervision in generative post-training through the lens of OT. Experiments on both short- and long-video generation show that $\texttt{PISCES}$ outperforms both annotation-based and annotation-free methods on VBench across Quality and Semantic scores, with human preference studies further validating its effectiveness. We show that the Dual OT-aligned Rewards module is compatible with multiple optimization paradigms, including direct backpropagation and reinforcement learning fine-tuning. Project page: https://roar-ai.github.io/pisces

Figures

Figures reproduced from arXiv: 2602.01624 by Cheng Zhao, David Gu, Dimitris Samaras, Gaurav Mittal, Mei Chen, Minh-Quan Le.

Figure 1
Figure 1. Figure 1: (a) Baseline (T2V-Turbo-v2) defines rewards over pre-trained VLM text-video embeddings, which suffer from distributional misalignment. (b) PISCES T2V post-training addresses this by formulating reward supervision over an OT-aligned embedding space. We propose a novel Dual OT-aligned Rewards module that aligns text embeddings to the video space, enabling effective visual and semantic alignment. (c) Compared… view at source ↗
Figure 2
Figure 2. Figure 2: PISCES T2V Post-Training. We introduce a Dual OT-aligned Rewards module: (i) a distributional OT map T ⋆ for Quality Reward via [CLS] representation similarity, and (ii) a discrete OT plan P ⋆ with spatio-temporal constraints for Semantic Reward via a Video-Text Matching (VTM) classifier. The rewards module provides supervision for fine-tuning the T2V denoiser and is applicable with direct backpropagation … view at source ↗
Figure 3
Figure 3. Figure 3: Human preference study. PISCES outperforms Hun￾yuanVideo, T2V-Turbo-v2, VideoReward-DPO in visual quality, motion and semantic alignment, validating its effectiveness in T2V. free approaches, achieving the highest scores across all metrics – Total, Quality, and Semantic [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of T2V models. PISCES produces videos with better semantic fidelity and visual quality, accurately capturing key details such as the reflective wet pavement and vibrant neon lighting [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cross-attention maps (left) are diffuse, OT plan without spatio-temporal constraints (middle) misaligns tokens, while our constrained OT plan (right) produces accurate correspondences. ward. Using only the Quality Reward (Row 3) improves the VBench Quality Score from 82.20 to 83.77 (Row 1), demonstrating its effectiveness in enhancing global coher￾ence and visual quality. In contrast, using only the Semant… view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE shows OT aligns text embeddings (green) closer to video embeddings distribution (orange). Pairwise distance dis￾tribution indicates OT preserves text embedding structure. A robot reaching for a red apple on the table. Rewards w/ Optimal Transport A robot reaching for a green apple on the table. Rewards w/ L2 Loss Mapping [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Post-training with OT-aligned rewards (left) yields con￾sistent outputs with expected color changes, while L2 loss mapping (right) causes sampling inconsistencies and visual artifacts. controlled experiment using the prompt “A robot reaching for a red/green apple on the table” with the same random seed. Post-training with OT-aligned rewards (left) main￾tains structural consistency – background, robot appea… view at source ↗
Figure 8
Figure 8. Figure 8: VTM accuracy under (γ, η) sweep. Despite configura￾tion variations, the accuracy remains stable (±1.48% std). Cross-Attention Map Our OT Plan 𝐏⋆ Mass m = 0.9 (✓ “glasses” token) Our OT Plan 𝐏⋆ Mass m = 0.5 (✗ ”glasses” token) “glasses” Prompt: The rabbit reads the newspaper and wears glasses [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Effect of OT mass parameter. Low OT mass (m = 0.5) fails to retain the “glasses” token, while higher mass (m = 0.9) improves alignment. partial mass m = 0.5 suppresses noise but also removes the valid token “glasses”, reproducing misalignment. Right: With mass m = 0.9, the OT plan retains the “glasses” token and aligns it better–but still imperfectly, activating only the right lens. This reveals a current … view at source ↗
Figure 10
Figure 10. Figure 10: Reward and gradient trends in post-training. N. Dual Rewards Interaction To assess the interaction between the OT-aligned quality and semantic rewards during training, we plot their reward trajectories and the cosine similarity of their gradients over time in [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

8 extracted references · 4 linked inside Pith

  1. [5]

    Ma, Y ., Wu, X., Sun, K., and Li, H

    URL https://openreview.net/forum? id=LyJi5ugyJx. Ma, Y ., Wu, X., Sun, K., and Li, H. Hpsv3: Towards wide- spectrum human preference score. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 15086–15095, October 2025. Pika Labs. Pika 1.0: Video generation model. https:// pika.art, 2023. Platform announcement; proprietary...

    Pith/arXiv arXiv 2025

  2. [6]

    Song, Y ., Dhariwal, P., Chen, M., and Sutskever, I

    URL https://openreview.net/forum? id=X8uQ1TslUc. Song, Y ., Dhariwal, P., Chen, M., and Sutskever, I. Consistency models. InICML, 2023. URL https://proceedings.mlr.press/v202/ song23a.html. Sun, X., Chen, Y ., Huang, Y ., Xie, R., Zhu, J., Zhang, K., Li, S., Yang, Z., Han, J., Shu, X., Bu, J., Chen, Z., Huang, X., Lian, F., Yang, S., Yan, J., Zeng, Y ., R...

    Pith/arXiv arXiv 2023

  3. [7]

    mechanism used to efficiently integrate OT-aligned re- wards into the PISCES post-training process. Section B de- scribes the Neural Optimal Transport (NOT) (Villani, 2009; Korotin et al., 2023) formulation for aligning the distribu- tions of text and video embeddings, while Section C presents our discrete token-level OT optimization via the entropic unba...

    2009

  4. [8]

    glasses” token)Our OT Plan 𝐏⋆Mass m = 0.5(✗

    domain, we curated 100 diverse out-of-distribution (OOD) prompts. These prompts cover challenging and un- derrepresented scenarios including robotics actions, embod- ied tasks, procedural instructions, abstract concepts, and rare object-event compositions. Example prompts include: • A robot arm with a red gripper picks up a blue cube and sorts it into a g...

    2024

  5. [2013]

    cc/paper_files/paper/2013/file/ af21d0c97db2e27e13572cbf59eb343d-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2013/file/ af21d0c97db2e27e13572cbf59eb343d-Paper. pdf. DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., Zhang, X., Yu, X., Wu, Y ., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., D...

    Pith/arXiv arXiv 2013

  6. [2023]

    URL https://openreview.net/forum? id=G5RwHpBUv0. KlingAI. Klingai - ai-powered solutions, 2025. URL https://www.klingai.com. Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., Wu, K., Lin, Q., Yuan, J., Long, Y ., Wang, A., Wang, A., Li, C., Huang, D., Yang, F., Tan, H., Wang, H., Song, J., Bai, J., Wu, J., Xu...

    Pith/arXiv arXiv 2025

  7. [2024]

    Izquierdo, S

    URL https://proceedings.mlr.press/ v235/huh24a.html. Izquierdo, S. and Civera, J. Optimal transport aggregation for visual place recognition. InProceedings of the ieee/cvf conference on computer vision and pattern recognition, pp. 17658–17668, 2024. Katageri, S., De, A., Devaguptapu, C., Prasad, V ., Sharma, C., and Kaul, M. Synergizing contrastive learni...

    2024

  8. [2025]

    Li, J., Feng, W., Fu, T.-J., Wang, X., Basu, S., Chen, W., and Wang, W

    URL https://openreview.net/forum? id=6kPBThI6ZJ. Li, J., Feng, W., Fu, T.-J., Wang, X., Basu, S., Chen, W., and Wang, W. Y . T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. InAdvances in Neural Information Processing Systems, 2024. Li, J., Long, Q., Zheng, J., Gao, X., Piramuthu, R., Chen, W., and Wang, W....

    2024