Predict the endpoint
A finite-difference endpoint estimate looks ahead to where each trajectory is heading.
FLOW MATCHING · INFERENCE-TIME CONTROL
Letting trajectories
spread.
More possibilities.
The same pursuit of quality.
Quality-Preserving Control for Diverse Flow Matching
Accepted atForty-Third International Conference on Machine Learning
Meet OSCAR. A training-free control mechanism that encourages text-to-image trajectories to explore complementary semantics, with diversity guidance orthogonal to the model’s flow.
FAMILIAR CONCEPTS. NEW POSSIBILITIES.Selected OSCAR generations from the authors’ result gallery ↗
¹ UNC Chapel Hill ² National University of Singapore ³ A*STAR CFAR
⁴ A*STAR IHPC ⁵ UC Santa Barbara ⁶ Nanyang Technological University
* Equal contribution † Corresponding author
01 / THE IDEA
Stochasticity
is not enough.
Preserving a trajectory’s marginal distribution does not guarantee an informative group of samples.
Our starting point was a question raised by reinforcement learning for visual generation. Methods such as Flow-GRPO and DanceGRPO brought group-relative optimization to generative models, using stochastic sampling to support exploration. Flow-GRPO, in particular, constructs an SDE that preserves the original flow’s time marginals. But distribution-preserving stochasticity leaves another question open: how much meaningful diversity does a finite group actually contain?
While experimenting with Flow-GRPO, we noticed that different rollouts could still produce semantically similar candidates. This matters for group-relative learning: equal rewards provide no reward-based preference signal, while small reward differences between redundant candidates may offer little reliable evidence about alternative generation strategies.
This observation motivated the question behind OSCAR: can we make a fixed set of trajectories explore complementary semantic outcomes while protecting generation quality? Rather than relying on noise alone, OSCAR introduces set-level feature-space guidance and orthogonal control to encourage trajectories to spread.
The RL observation motivated this work; the experiments presented here evaluate generation diversity, quality and alignment, rather than downstream RL training gains.
+15.1%
Semantic diversity · Vendi (Inception)
+20.7%
Low-level diversity · Vendi (Pixel)
−11.8%
Distribution discrepancy · KID
0
Additional training required
Relative changes vs. FM-SD3.5 on T2I-CompBench spatial color / complex subsets (paper, Table 2). These gains describe this evaluation setting, not every prompt or backbone.
02 / SEE THE DIFFERENCE
Same concept.
Wider horizons.
Explore real generated image sets.
Switch concepts and compare methods side by side.
CADSBASELINE
OSCAROURS ↗
Selected qualitative examples from the supplied results repository. Columns show method-specific sample sets; image positions do not establish matched seeds. Click an image to inspect it.
03 / UNDER THE HOOD
Go forward.
Explore sideways.
A geometric safeguard for diversity:
control acts orthogonally to the base velocity.
OSCAR · spreading trajectories
NOISECONCEPTUAL ILLUSTRATION · NOT RECORDED MODEL TRAJECTORIESSEMANTIC MODES
Spread in feature space
A set-level feature-volume objective encourages complementary semantic outcomes.
Protect the flow
Orthogonal projection and redundancy-aware reweighting constrain the diversity control; scheduled noise adds exploration.
Read the abstract +
Flow-based text-to-image models follow deterministic trajectories, making it costly to explore diverse modes under limited sampling budgets. We present a training-free, inference-time control mechanism that makes the flow itself diversity-aware. Guidance encourages lateral spread in feature space and introduces time-scheduled stochastic perturbations orthogonal to the generation flow. Across the paper’s text-to-image settings, OSCAR improves set-level diversity while maintaining competitive image quality and prompt alignment.
Condensed from the supplied paper. See the PDF for the complete abstract, theoretical assumptions and experimental protocol.
04 / ACROSS SAMPLING BUDGETS
Diverse results.
Across step budgets.
Explore OSCAR outputs at different sampling budgets.
Drag the slider to compare 10, 20 and 40 steps.
Selected OSCAR results at three sampling budgets. Each slider position shows outputs generated with the selected number of steps, rather than intermediate frames of a single generation.
05 / MEASURED, NOT JUST SEEN
The numbers
behind the spread.
Inspect diversity, image quality and alignment.
All values below are transcribed from the paper.
Bar values show means; uncertainty is preserved in the table.
What do these metrics measure? +
Vendi ↑ Effective diversity in a sample set. Pixel captures low-level variation; Inception captures feature-space diversity.
1 − MS-SSIM ↑ Structural dissimilarity between generated samples.
FID / KID ↓ Discrepancy between generated and reference feature distributions. Compare within the same evaluation protocol.
CLIP Score ↑ Text–image alignment. CLIP-IQA ↑ Learned image-quality assessment.
BRISQUE ↓ No-reference image-quality assessment. Image Reward ↑ Learned human-preference alignment.
DIM / CIM Default attribute balance (DIM closer to zero) and explicit attribute controllability (CIM higher).
06 / BEYOND THE AVERAGE
More modes.
A broader view.
Original experimental figures.
Open any figure for a closer look.
07 / COMPOSITION & DETAIL
Beyond a
single concept.
Color, spatial relationships, complex scenes
and challenging human portraits.
PORTRAIT / MATERIAL / LIGHTINGOpen full resolution ↗
Selected qualitative comparisons. Method labels are preserved: OSCAR, PG, CADS, DPP, Base Model, Mix ODE/SDE and APG. On smaller screens, scroll horizontally to inspect each method.
08 / TAKE IT FURTHER
Let your
trajectories spread.
01 / PAPERRead the full story ↗02 / IMPLEMENTATIONExplore the code ↗03 / EXPERIMENTSBrowse the result archive ↗
Scope & limitations +
OSCAR operates on jointly active sets of trajectories and is not designed for a strict one-image-per-prompt setting. Endpoint feature extraction and diversity-gradient computation add inference overhead relative to the original sampler. Results depend on the model, prompts and evaluation setting; see the paper’s limitations and cross-backbone results.
Cite this work
@article{wu2026letting,
title={Letting Trajectories Spread: Quality-Preserving Control for Diverse Flow Matching},
author={Wu, Jingxuan and Wan, Zhenglin and Yu, Xingrui and Yang, Yuzhe and An, Bo and Tsang, Ivor and You, Yang},
year={2026}
}Title, authors and year follow the supplied manuscript. Conference acceptance is listed above as provided by the authors; no DOI is inferred.