- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes a training-free method for improving diversity in flow-matching models by reversing the velocity field during early sampling steps. The paper designs a ease-out schedule for early exploration and gradually restores the original velocity for later refinement. The method is motivated by the relationship between velocity divergence and differential entropy and requires no additional network evaluations or interactions between samples. Experiments show promising diversity-quality trade-offs, including evaluations using minimum quality per prompt. Additional experiments demonstrate its effectiveness on distilled image models, video generation, 3D generation, and unconditional image generation.
Critical strengths
- The proposed intervention method is simple and practical. It modifies only the predicted velocity, preserves independent sampling, and introduces negligible additional computation.
- The evaluation on image generation is comprehensive, covering multiple baseline families and several diversity and quality metrics. Reporting minimum quality and mean quality is particularly useful for checking whether diversity gains come from a few poor samples.
- The experiments provide promising evidence that velocity inversion changes meaningful image structure. The ablation studies and evaluation complement the qualitative results and support the proposed method.
Critical weaknesses
- The theoretical analysis does not fully justify the claimed improvement in mode coverage. In Equation (10), the assumption that the expected divergence under the corrected sampling distribution is negative, may not remain valid throughout the entire sampling trajectory. And, an increase in differential entropy does not necessarily imply improved mode coverage, as it could instead come from greater dispersion within existing modes.
- The ablation studies are insufficient to clearly isolate the contribution of velocity inversion. The proposed approach simultaneously modifies the direction and magnitude of the velocity and the cumulative displacement along the trajectory. Conducting more rigorous comparative experiments, e.g. controlling for perturbation intensity or output quality, would help determine whether velocity inversion itself is the key factor driving the observed performance improvements.
- The evaluation mainly relies on feature-based diversity metrics, which cannot clearly distinguish the recovery of rare semantic modes from variations in viewpoint, composition, or appearance. More explicit evaluations of semantic mode coverage on real-world conditional generation tasks would strengthen the results.
Questions for authors
- Can the authors report divergence statistics across prompts under the modified sampling trajectories, including how often the negative-divergence assumption fails during the velocity inversion interval?
- Can the authors compare the proposed method against non-negative time-dependent velocity schedules with comparable perturbation strengths or output quality?
- How sensitive is the proposed method to the number of sampling steps? Does the diversity–quality trade-off remain consistent when varying the sampling budget?
3 comments
Response to Reviewer PbLs: W1, W2, Q2, Q3
W1 / W2 / Q2. The theoretical analysis does not fully justify the claimed improvement in mode coverages; non-negative schedules at matched strength or quality.
We agree that increased entropy alone does not imply diversity, as adding noise also increases entropy. We also acknowledge that our theory does not fully explain mode coverage. A guarantee would require characterizing how the learned velocity field partitions the noise space into modes, which is beyond current analysis. We therefore rely on the empirical evidence below.
To investigate the contribution of velocity inversion, we compare against the following schedules:
- Our exponential function but with η = 1: , which never goes below 0.
- Ramps: , with
- linear:
- cosine:
- exponential:
For each η of our main sweep (1.5, 2.0 and 2.5; Sec. 5), we set σ or w so that each schedule’s perturbation strength, measured as the average of over all steps, matches ours. All runs use FLUX.1-dev, 20 steps, 10 images per prompt and the same starting noise, on 50 GenEval test prompts.
Table R4. All rows within a block have the same perturbation strength (mean α, column 2). All metrics are higher-is-better; ▼ / ▲ = statistically significantly lower / higher than ours in the same block (two-sided paired t-test over prompts, p < 0.05).
| Method | Mean α | CLIP ↑ | HPSv2 ↑ | DINO ↑ | DreamSim ↑ |
|---|---|---|---|---|---|
| Matched strength: mean α = 0.61 (ours at η = 1.5) | |||||
| Ours, η = 1.5 | 0.611 | 28.09 | 28.97 | .705 | .425 |
| Our function, η = 1, σ = 0.42 | 0.610 | 28.66▲ | 28.74▼ | .687▼ | .396▼ |
| Linear ramp, w = 0.77 | 0.609 | 28.11 | 29.07▲ | .698 | .409▼ |
| Cosine ramp, w = 0.77 | 0.609 | 27.99 | 29.00 | .704 | .413▼ |
| Exponential ramp, w = 0.48 | 0.608 | 27.76▼ | 28.81▼ | .721▲ | .443▲ |
| Matched strength: mean α = 0.48 (ours at η = 2.0) | |||||
| Ours, η = 2.0 | 0.481 | 27.10 | 28.18 | .753 | .503 |
| Our function, η = 1, σ = 0.66 | 0.480 | 27.66▲ | 27.79▼ | .706▼ | .375▼ |
| Linear ramp, w = 1.04 | 0.481 | 27.22 | 28.03 | .725▼ | .426▼ |
| Cosine ramp, w = 1.04 | 0.481 | 26.66 | 27.99▼ | .743 | .472▼ |
| Exponential ramp, w = 0.64 | 0.485 | 24.82▼ | 27.17▼ | .777▲ | .542▲ |
| Matched strength: mean α = 0.35 (ours at η = 2.5) | |||||
| Ours, η = 2.5 | 0.351 | 24.74 | 26.90 | .786 | .555 |
| Our function, η = 1, σ = 1.06 | 0.350 | 24.89 | 25.62▼ | .695▼ | .266▼ |
| Linear ramp, w = 1.42 | 0.352 | 24.80 | 25.78▼ | .717▼ | .308▼ |
| Cosine ramp, w = 1.34 | 0.351 | 24.47 | 25.57▼ | .750▼ | .401▼ |
| Exponential ramp, w = 0.81 | 0.354 | 20.09▼ | 23.54▼ | .736▼ | .475▼ |
At η = 1.5 and 2.0, ours is more diverse on DreamSim than every schedule except the exponential ramp, which gains diversity only by losing quality: at η = 2.0 its CLIP drops to 24.82 (ours 27.10), consistent with corrupted images. At η = 2.5, ours is more diverse than every schedule on DINO and DreamSim and has higher HPSv2.
Q3 Is the trade-off consistent across budgets? Sampling step sensitivity?
The diversity–quality trade-off is consistent across budgets in terms of rank order: at every N total number of sampling steps, a larger η gives higher diversity (DINO and DreamSim: η = 1.5 < 2.0 < 2.5) and lower CLIP (η = 1.5 > 2.0 > 2.5); only the values shift slightly. Setup: we reran Base and ours (η = 1.5, 2, 2.5) with N = 15, 20, 25 and 30 steps on 50 validation prompts. Note that the schedule is defined on the normalized step index , so the inverted window ends at nearly the same noise level for every N.
Table R4. Mean ± SE over prompts. Quality: CLIP; diversity: DINO, DreamSim.
| N = 15 | N = 20 | N = 25 | N = 30 | ||
|---|---|---|---|---|---|
| (Quality) CLIP ↑ | Base | 26.88±.37 | 26.82±.37 | 26.77±.39 | 26.66±.39 |
| η=1.5 | 26.84±.37 | 26.90±.37 | 27.05±.36 | 27.08±.37 | |
| η=2.0 | 26.02±.35 | 26.30±.36 | 26.36±.37 | 26.49±.38 | |
| η=2.5 | 24.66±.32 | 25.02±.33 | 25.13±.36 | 25.07±.34 | |
| (Diversity) DINO ↑ | Base | .639±.009 | .642±.009 | .642±.009 | .645±.009 |
| η=1.5 | .729±.007 | .727±.007 | .729±.007 | .729±.007 | |
| η=2.0 | .759±.006 | .754±.006 | .754±.006 | .755±.006 | |
| η=2.5 | .776±.004 | .770±.004 | .770±.004 | .770±.004 | |
| (Diversity) DreamSim ↑ | Base | .284±.013 | .284±.013 | .286±.013 | .289±.013 |
| η=1.5 | .425±.013 | .416±.013 | .415±.013 | .413±.013 | |
| η=2.0 | .511±.010 | .494±.010 | .492±.010 | .489±.010 | |
| η=2.5 | .553±.007 | .541±.007 | .538±.007 | .535±.007 |
Response to Reviewer PbLs: W3
W3. Feature-based diversity metrics cannot separate rare semantic modes; more explicit evaluation of semantic mode coverage is required.
We thank the reviewer for the suggestion.
Setup. To measure whether a method reaches different semantic modes, we follow GRADE [Rassin et al., 2024] by using a VLM to detect predefined modes in each image and measuring their spread with the normalized entropy Ĥ.
We use 50 underspecified prompts specifying various objects / scenes (e.g., “a photo of a vase”, “a photo of a fox”, …) and generate 10 images for each prompt. To evaluate images from a given prompt (e.g., "..frog"), we first predefine five kinds of attributes (e.g., "color", "action", "kind"), each with 10 possible values, including or “none of these / cannot tell”. These attributes and their values are different for each prompt and are designed by Claude Opus 5.5. During evaluation, we use Qwen2-VL-7B to classify each image into one of the 10 values for each attribute. We measure mode coverage as the normalized entropy Ĥ of these classes, and report CLIP score (text alignment) and HPSv2 (image quality).
We compare against representatives of each baseline type: IG, CNO, SPELL, with matched quality: For each baseline, we tuned their hyperparameters using a separate validation set (also 50 prompts x 10 images) and picked the ones whose CLIP and HPSv2 scores are at most 1% below those of Base (one-sided paired t-test, p < 0.05).
Table R5. Mean ± SE over prompts. ▼ / ▲ = statistically significantly lower / higher than ours. All marks: two-sided paired t-test over prompts, p < 0.05.
| Method | CLIP ↑ | HPSv2 ↑ | Ĥ ↑ |
|---|---|---|---|
| Base, cfg 3.5 | 24.64 ± 0.24 | 28.70 ± 0.12 | 0.242 ± 0.014 ▼ |
| IG [0.01, 0.99] | 24.59 ± 0.26 | 28.70 ± 0.12 | 0.267 ± 0.013 ▼ |
| CNO, γ = 1 | 24.62 ± 0.27 | 28.65 ± 0.12 | 0.268 ± 0.016 ▼ |
| SPELL, r = 150 | 24.53 ± 0.26 | 28.61 ± 0.14 | 0.248 ± 0.015 ▼ |
| Ours, η = 1.25 | 24.58 ± 0.26 | 28.70 ± 0.13 | 0.330 ± 0.014 |
Summary. Ours has a statistically significantly higher Ĥ than every other baselines. Note that the test CLIP and HPSv2 scores of all baselines remain comparable when they were tuned for matched quality on the validation set.
Qualitatively, examples of rare semantic modes only our method produced include:
- Object categories: a saxophone or a trumpet for “a photo of a musical instrument” (normal: guitar, piano, violin);
- Species: a dragonfly for “a photo of an insect” (normal: butterfly, bee, ladybug), and a horned viper for “a photo of a snake” (normal: python, rattlesnake, garter snake);
- Dishes: waffles for “a photo of a breakfast” (normal: eggs and bacon, pancakes, cereal with milk);
- Subjects: a mythical creature for “a photo of a sculpture” (normal: human figure, bust or head, abstract shape).
Response to Reviewer PbLs: W1, Q1
W1 / Q1. Does the negative-divergence assumption hold along the trajectory? Divergence statistics across prompts and failure rate in the inversion interval.
In our experiment, E[∇⋅vθ]<0 at every step, and ∇⋅vθ<0 in all 1,000 / 2,000 / 2,500 inverted (trajectory, step) pairs for η = 1.5 / 2 / 2.5. The experiment estimates the divergence of the unscaled velocity with Hutchinson’s estimator (4 Gaussian probes per step) on 500 trajectories (50 T2I-CompBench validation prompts × 10 seeds; FLUX.1-dev, 20 steps). The entropy estimate rises at every inverted step and falls once αk>0, while for Base it decreases monotonically (Table R3).
Table R3. Per-step statistics over the 500 trajectories, mean ± SE over prompts: (a) E[∇⋅v] of the unscaled velocity; (b) entropy change Hk−H0 (nats). Bold = inverted steps (αk<0).
(a) E[∇⋅v], where v=vθ(xt,t) is the model’s velocity before our scaling, not αkv.
(b) Hk−H0 (nats)