Abstract
A novel on-policy generative field distillation framework called DanceOPD is proposed to unify text-to-image generation, local editing, and global editing capabilities in flow-matching models through capability-specific routing and velocity-based training.
Modern image generation demands a single model that unifies diverse capabilities, including text-to-image (T2I), local editing, and global editing. However, these capabilities are rarely naturally aligned and often conflict. For instance, editing tends to degrade T2I performance, while global and local editing interfere with each other. Consequently, effectively composing these capabilities has become a central challenge for image generation model training. To tackle this, we introduce DanceOPD, an on-policy generative field distillation framework for flow-matching models that routes each sample to one capability field, queries one low-noise student-induced state, and trains with a simple velocity MSE objective. With each capability source defined as a velocity field over the shared flow state space, the student learns from fields queried on its own rollout states to compose expert capabilities. This formulation also absorbs operator-defined fields such as classifier-free guidance. Comprehensive experiments on T2I, editing, realism-field absorption, and CFG absorption show that our approach improves multi-capability composition, strengthening target capabilities while preserving anchor generation quality. We believe this work establishes a practical route for generative field distillation in flow-matching models.
Community
DanceOPD uses hard routing to select one frozen capability field, queries it on a low-noise on-policy student state, and matches the selected velocity with a local velocity MSE loss.
Image-generation post-training is increasingly becoming a process of training specialized experts and absorbing their capabilities into one unified model. T2I, local editing, global editing, style control, realism enhancement, and CFG-like guidance are all desirable in the same checkpoint. Yet in practice, data mixing and weight merging often cause interference: new skills improve, while existing ones degrade.
DanceOPD addresses this problem through on-policy field distillation for flow-matching models. We view each expert capability as a velocity field in a shared latent flow space. The student rolls out its own trajectory, queries the corresponding teacher at a low-noise semantic state, and learns with a simple velocity-MSE loss. Remarkably, a single teacher query is sufficient for effective capability absorption.
On T2I + Edit composition, DanceOPD reaches 5.347 on GEditBench, outperforming the best reproduced OPD baseline by 8.1%, while maintaining 0.849 on GenEval. On the more conflicting Local + Global Edit setting, it achieves 5.498 on GEditBench, improving over the strongest baseline by 16.1%, with 0.848 GenEval. The same framework also supports realism absorption and CFG absorption.
Compared with recent DiffusionOPD and FlowOPD, our focus is not merging RL-trained models from different reward models, but studying how generative capabilities themselves can be stably absorbed and composed in a shared flow space. We further analyze teacher routing, query states, KL-style objectives, and dense vs. single-query designs, showing about 10ร training efficiency over DiffusionOPD.
We hope DanceOPD offers a cleaner path for visual-generation post-training: not forcing data or weights together, but letting the model learn expert capabilities on its own distribution.
๐ Paper: https://arxiv.org/abs/2606.27377
๐ Project: https://danceopd.github.io/
Please consider upvoting ๐
๐ ๏ธ Code is still under approval; pseudocode is provided in the paper.
Get this paper in your agent:
hf papers read 2606.27377 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper