acceptodds
Under review as a conference paper at ICLR 2027

Grid-based Repulsive Sampler for Distribution Rebalancing in Generative Post-Training

Abstract

Post-training a generative model begins by reducing a large candidate pool to a small fine-tuning set. Near-duplicates are removed above a similarity threshold and over-represented regions are capped, once, before training, and neither decision can be revisited afterwards. We propose the Grid-based Repulsive Sampler (GridReS), which rebalances the training distribution and keeps the pool intact. GridReS constructs a uniform grid over a low-dimensional projection of a pretrained embedding space. During training, occupancy-aware sampling redistributes probability mass across cells, while intra-cell repulsion scores candidates against a memory of recent draws from the surrounding neighbourhood. The partition supplies the locality that keeps repulsion at a bounded per-draw cost; repulsion controls the redundancy inside a region, which weighting over regions cannot reach. At coverage identical to random sampling, GridReS cuts cross-source near-duplicate co-exposure by 50% on the fine-tuning pool of Kandinsky-6 Lite, a text-to-audio-video model, and by 25% on OpenVid-100k. When Kandinsky-6 Lite is fine-tuned as a single job on the full pool, GridReS improves VBench imaging quality over random sampling by 1.84 points, beyond run-to-run variance, and trained per domain it matches the deployed deduplication pipeline at equal optimizer steps without discarding any data. Code is available at https://github.com/94yvwj2hrj-eng/gridres.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
distribution rebalancing, data curation, balanced sampling, data coverage, post-training, supervised fine-tuning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_K819OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GridReS is an online sampler for video post-training that replaces offline deduplication. Clips are bucketed on a PCA grid over embeddings, cells are drawn ∝ n^α, and within a cell the sampler prefers clips far from recent draws. At equal coverage it halves near-duplicate co-exposure R on the Kandinsky-6 pool and lowers it by 25% on OpenVid-100k. A single fine-tuning job gains 1.84 VBench imaging points over random sampling, and per-domain jobs match production deduplication.

Critical strengths
  1. The paper reports its seed floor and negative results very openly. The analysis in App. B of what region-level samplers can and cannot do is useful on its own.
  2. Per-draw cost is low, and the code is released.
Critical weaknesses
  1. The link from R to model quality is not shown. Per domain, random sampling over the full pool also matches production (Table 11), so that result doesn't isolate GridReS. In the single job, GridReS covers only 19.6% of the pool vs. 33.7% for random and is worse on V-JEPA 2 metrics, so the imaging gain may come from which clips were seen. The evaluation also lacks text alignment, audio metrics and human preference.
  2. E is precomputed and the sampler uses no training signal, so the natural baseline is a simple offline permutation that spreads near-duplicate groups apart. It is missing.
Questions for authors
  1. How does an offline permutation that spreads near-duplicate groups compare?
  2. In the single-job run, can you separate the effect of which clips were selected from the effect of their order?
  3. How do R and coverage behave with more ranks?
Reviewer_jL9vOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GridReS, an online data sampler for post-training video generation models that aims to replace offline deduplication and per-region capping. It partitions a pretrained embedding space with a uniform grid, reweights cells by occupancy, and within each cell prefers clips that are dissimilar to recent draws. The authors report reduced near-duplicate co-exposure in simulation and comparable or improved VBench scores compared to random sampling and a production curation pipeline, without discarding any data.

Critical strengths
  1. The question is practical: whether curation thresholds can be deferred to sampling time.
Critical weaknesses
  1. This article claims to have improved the quality of video generation, but it does not show any of the generation results. No supplementary videos, and no failure cases. The claimed quality gain is not supported, quantitatively or qualitatively.
  2. There are no side-by-side comparisons of GridReS and random under the same prompt and the same seed.
  3. Readers are completely unable to determine whether a MUSIQ score of +1.84 corresponds to a visible difference to the naked eye or is merely an indicator noise.
  4. The method details in figure 2 is like a thumbnail of the slide rather than a method diagram. Although it contains a lot of texts and plots, it is hard for the audience to understand the sampling procedure: most of the text is truncated fragments with no connection between them (e.g. “drawn ∝ s”, “computed once, cached”), the links between panels are labelled only “score” and “push into Q”, and results and implementation details (Panels A, F, G) are mixed into the method.
  5. RR does not connect to the claimed benefit. R≡1R \equiv 1 at whole passes, so it only measures within-pass deferral, and most per-domain jobs exceed a pass. Where RR differs most, quality ties both production curation and full-pool random sampling. No experiment shows that keeping the discarded data helps.
  6. Hyperparameters are selected on the evaluation set, and the configurations of each experiment are inconsistent.
Questions for authors
  1. What is the side-by-side generations from GridReS and random sampling (single job) for the same prompts and seeds? Show some supplementary videos, including failure cases. Is the +1.84 MUSIQ difference visible to the naked eye?
  2. Can you add single-job controls with at least three seeds each: grid-only at α=0.5\alpha=0.5, GridReS at α=1\alpha=1, and a coverage-matched run? Without them, the gain cannot be separated from cell reweighting, lower coverage or seed variance. Also, how do you reconcile the VBench gain with the loss on all four V-JEPA 2 metrics?
  3. Is there any experiment showing that keeping the discarded data helps, for example on long-tail prompts or generation diversity? Can you report a windowed RR on the multi-pass per-domain logs?
  4. Were α\alpha, target occupancy and dd selected on held-out data? If not, how do results change on a held-out pool with the hyperparameters fixed? Can the OpenVid-100k models be retrained with the configuration of Section 2?
  5. How does GridReS differ from temperature sampling combined with embedding-based diversification (e.g. D4) or online batch selection (e.g. JEST)? How does it compare to repulsion applied directly on the precomputed graph EE, which needs no grid?
  6. Please redraw Figure 2.
Reviewer_MPSgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GridReS, an online data sampler meant to replace the two offline curation steps in video-generation post-training: near-duplicate removal and per-region capping. Each clip is embedded once with Qwen3-VL-Embedding-8B (R4096\mathbb{R}^{4096}). The embeddings are PCA-projected to d=4d=4 and cut by a uniform grid whose side ℓ\ell is bisected to a target mean occupancy. At each draw, the sampler does two things:

  1. Cell weighting. It picks a cell with probability ∝ncα\propto n_c^\alpha.
  2. Intra-cell repulsion. It picks an unserved member of that cell with probability proportional to its full-space cosine distance to the nearest entry of a FIFO queue (M=128M=128). The queue is shared across the cell's 3d3^d neighbourhood. Sampling is without replacement, so each rank's stream is a permutation of the pool per pass.

Redundancy is measured by R(p)=2∑(i,j)∈EpipjR(p)=2\sum_{(i,j)\in E}p_ip_j, the near-duplicate co-exposure over a cross-source cosine >0.9>0.9 graph, reported as a ratio to random sampling.

The main claims, at coverage matched to random sampling:

  • Co-exposure. RR drops by about 50% on the private Kandinsky-6 SFT pool (67k clips) and by about 25% on OpenVid-100k.
  • Single-job training. Fine-tuning Kandinsky-6 Lite (3B) as a single job at 0.60 epochs improves VBench imaging quality by +1.84 over random sampling, beyond a seed floor.
  • Per-domain training. Trained as 11 per-domain souped jobs, the undeduplicated pool matches the production dedup pipeline.

Appendices give a perturbation argument for why region-level reweighting alone trades redundancy for coverage (Eqs. 3, 5, 6, 8), plus encoder and projection analyses and an extensive set of ablations and negative results.

Critical strengths
  1. A clean argument for acting within regions, with an honest decomposition of the mechanism.

    • Appendix B.3 shows that any region-level sampler is bound by a redundancy-versus-coverage trade: scaling a cell's contribution to RinR_{in} by λ\lambda scales its coverage by λ\sqrt{\lambda}. A within-cell policy can lower RR at fixed coverage.
    • The ablations then confirm the consequence empirically. Occupancy weighting is a weak density proxy (correlation +0.11+0.11) and raises RR on OpenVid for α<1\alpha<1, while repulsion lowers RR at every α\alpha on both pools (Tables 13–14).
    • This reframes the grid as a locality structure, essentially LSH, rather than a density model. That framing is the most transferable insight of the paper.
  2. Unusually careful and self-critical empirical reporting.

    • VBench differences are read against a retrain seed floor, not just prompt-bootstrap intervals.
    • Zero-RR runs are dropped and counted rather than clipped (H.6 shows clipping would invert a ranking).
    • Distribution metrics are evaluated in V-JEPA 2, deliberately chosen as the space least aligned with the sampler's own.
    • Negative results are reported prominently (Sec. 6).

    This makes the claims auditable and raises my trust in the numbers that do survive.

Critical weaknesses
  1. The downstream quality claim is thin and internally contested.

    Per-domain schedule. This is the production-relevant setting, and here GridReS only matches production curation, with no dimension beyond the seed floor. Against random sampling on the full pool it trails on subject and background consistency.

    Single job. The headline +1.84 imaging gain rests on one run, on one of five VBench dimensions, and it conflicts with other evidence from the same run:

    • It trails its random control in V-JEPA 2 precision, density and coverage beyond the seed gap (Table 12).
    • Its draw log covers only 19.6% of the pool versus 33.7% for random, because the eight independent rank streams concentrate on the same sparse cells (H.7).

    So in the one setting where the model "wins", the sampler is not at matched coverage. H.7 admits that the coverage-adjusted nulls do not separate. The live sampler therefore behaves partly like an implicit subset selector, and its distributed design (independent per-rank queues, no cross-rank repulsion) is a material flaw rather than a detail.

    What is left. The case that lower RR improves the generative model reduces to "no measurable cost". That is a useful but much weaker contribution than the abstract implies. No evaluation covers text alignment, audio (for a T2AV model), or memorization, and memorization is the obvious risk of keeping near-duplicates.

  2. The metric and the missing simple baseline leave the necessity of the online mechanism unproven.

    What RR measures. By construction R≡1R\equiv 1 at every whole pass, so RR only measures the spacing of near-duplicates within a pass. Six of the 11 per-domain jobs exceed one pass, and no windowed variant is computed (acknowledged in Sec. 6).

    The missing baseline. The paper already computes the exact near-duplicate graph EE offline (12,737 edges). Given EE, a trivial offline permutation is available: spread each connected component's members evenly through the epoch, for example by stratified or round-robin interleaving of components. That permutation directly maximizes within-pass spacing at zero online cost and identical coverage, with no grid, PCA, queue, or per-rank state.

    Why it matters. Without this baseline, and without a windowed-RR or model-level comparison against it, it is unclear what the online, approximate, grid-local repulsion buys. That matters especially because RR is defined in the same space and graph the sampler optimizes, which makes the distributional result partially circular.

    Novelty. Relative to DPP mini-batch diversification and LSH bucketing, the novelty is primarily in the systems integration.

Questions for authors
  1. Offline spacing baseline. How does an offline permutation built from EE compare on RR and VBench? For example, interleave members of each connected component at maximal spacing within a pass. If it matches GridReS, what is the case for the online sampler?
  2. Robustness of the +1.84 single-job gain.
    • Can you report a second seed of the single job (both arms)?
    • Can you explain the Table 12 reversal, where single-job GridReS is worse in precision, density and coverage?
    • Does a cross-rank fix restore matched coverage, and does the VBench gain survive it? Candidates include a shared queue, rank-offset cell schedules, or sharding the pass across ranks.
  3. Past one pass. Can you report a windowed RR (co-draws within the last WW draws) for the per-domain jobs that exceed one pass? This would test whether the spacing effect persists where RR is uninformative.
  4. Configuration mismatch. The trained OpenVid models and the per-domain (d=3, M=32) runs differ from Section 2 in score, memory, soft membership and with/without replacement. Which trained results correspond exactly to the released default configuration?

Suggestions (not required for my assessment):

  1. Measure memorization or extraction on the near-duplicate groups (e.g., the 175/152/142-clip groups), given the ethics statement's own concern.
  2. Test the EE-aligned cell weighting suggested by Eq. 5, which the paper notes but does not run, and report text-alignment and audio metrics.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.