DanceOPD: On-Policy Generative Field Distillation

Unifying Generative Fields with DanceOPD

Modern generative image modeling has split into multiple specialized subfields. Models are typically designed to solve a single primary problem well, such as creating new images from text descriptions, editing localized parts of existing images, or applying overall stylistic changes to a picture. While this specialization yields high quality within each task, deploying separate models for every task creates massive operational overhead and increases compute costs.

The obvious solution is to train a single model to handle all of these tasks. However, researchers have long faced a fundamental obstacle known as capability interference. When a single neural network is trained on multiple diverse tasks, the training directions for these tasks often conflict. This conflict causes the model's performance to degrade across all tasks, leaving it unable to produce high-quality text-to-image outputs or precise image edits.

To address this limitation, the DanceOPD framework introduces On-Policy Generative Field Distillation. By treating each specialized teacher model as a velocity field over a shared latent flow space, the student model learns to distill these multiple fields into a single, efficient student model. Through hard-routed training paths and on-policy supervision, DanceOPD achieves state-of-the-art results across image generation and editing tasks without capability degradation.

The Mechanics of Capability Interference

To understand why multi-task image models struggle, it is necessary to examine how neural networks receive optimization signals during training. In a standard multi-task setup, a network is presented with mixed training data: some batches focus on text-to-image generation, others on local image editing, and others on style transformations.

During backpropagation, the model calculates the gradient—the direction in parameter space that reduces error—for the current batch. For a text-to-image task, the gradient encourages the model to generate layout and structural details from random noise. In contrast, for a local editing task, the gradient instructs the model to keep the majority of the image layout identical, changing only a small, specific region.

When these tasks are combined, the resulting gradients conflict. The update step for one task overrides the updates for another. Over many training cycles, this conflict leads to a compromise in the parameter space where the model performs mediocrely at everything. It fails to generate sharp details from text prompts and struggles to maintain image consistency during edits, resulting in blurry details and poor instruction compliance.

This issue is compounded by the fact that different tasks require different levels of noise conditioning. Text-to-image generation typically operates across the entire noise spectrum, from high-variance starting points to low-variance outputs. Image editing, particularly local editing, relies on preserving low-frequency structures, which requires operating within restricted noise zones. When a single network attempts to map these varying paths, the optimization landscape becomes highly fragmented and unstable.

Hard-Routed Sample-Wise Field Matching

DanceOPD introduces Hard-Routed Sample-Wise Field Matching to resolve these gradient conflicts. Instead of averaging gradients across mixed tasks or attempting to force the student model to learn a blended trajectory, the framework routes each training sample to exactly one specialized teacher velocity field.

Mathematically, the training process treats the teachers as velocity fields over a shared latent flow. When a training sample is processed, the system assigns it to a specific task category, such as text-to-image synthesis. The student model is then optimized to match the velocity field of the corresponding teacher for that sample.

This hard routing ensures that the student parameters are updated in a clear, consistent direction for each sample. The model is never forced to reconcile conflicting task demands on a single update step. By isolating the trajectories, the student network learns to partition its internal representation space, developing specialized subnetworks or pathways that handle different tasks while sharing core features. This approach eliminates target-field ambiguity and preserves model capacity.

On-Policy Field Querying: Eliminating Covariate Shift

Standard distillation frameworks often rely on off-policy training, where the student model is trained on data trajectories pre-generated by the teacher. While this is simple to implement, it suffers from a significant limitation during deployment.

During inference, the student model generates its own trajectory step by step. Because the student is not a perfect replica of the teacher, it makes small errors at each step. In an off-policy setup, the student has only been trained on the teacher's clean paths. It does not know how to correct these errors, causing them to accumulate over successive steps. This covariate shift leads to distorted images and artifacts.

DanceOPD solves this by using On-Policy Field Querying. During training, the student model generates its own rollout states. The framework then queries the teacher velocity fields at these student-generated states rather than the teacher's states. This forces the student model to learn how to move toward the target from the actual states it encounters during generation.

By querying on-policy, the student receives feedback on its own errors. The teacher fields guide the student back to the correct path, training the model to be self-correcting. This results in greater stability during generation, producing sharper details and better adherence to input prompts even when using a small number of inference steps.

Semantic-Side Single Query Optimization

While on-policy distillation provides superior results, it introduces high computational overhead. In a naive implementation, querying the teacher model at every step of the student's rollout requires running the teacher's full forward pass multiple times per sample. This slows down training and requires massive memory bandwidth.

To make on-policy training practical, DanceOPD introduces the Semantic-Side Single Query method. This technique operates on the observation that the early stages of the latent flow matching path determine the high-level semantic layout of the image, while the later stages focus on low-level detail refinement.

Instead of querying the teacher at every single step, DanceOPD performs a single query on the semantic portion of the latent space. This single query provides the essential directional guidance needed to align the student's trajectory. By avoiding redundant queries during the detail-refinement phase, the framework significantly reduces computational overhead and minimizes path correlation. This optimization makes it possible to train DanceOPD models in a fraction of the time required by standard on-policy distillation frameworks.

Benchmark Performance: Outperforming Baselines

To verify the effectiveness of the framework, DanceOPD was evaluated on GEditBench, a benchmark designed to assess multi-task image models. The evaluation compared DanceOPD to several state-of-the-art distillation baselines, measuring both image quality and instruction alignment.

The results demonstrate that DanceOPD achieves significant improvements over competing methods:

As shown in the table, naive multi-task distillation suffers from capability interference, resulting in worse scores than the baseline student model. DanceOPD, however, achieves substantial improvements. Specifically, it delivers a 16.1% improvement in local-plus-global editing composition and an 8.1% improvement in the combination of text-to-image generation and editing.

Crucially, this is achieved while preserving the quality of the baseline text-to-image model. The anchor generation scores remain within 0.1% of the baseline, proving that DanceOPD unifies editing capabilities without degrading the model's core generation performance.

Integrating Classifier-Free Guidance

An additional benefit of the DanceOPD framework is its ability to absorb operator-defined fields during the distillation process. In standard image generation models, Classifier-Free Guidance (CFG) is used to control the trade-off between image quality and prompt alignment. However, CFG requires running two forward passes per step during inference, doubling the computational cost.

DanceOPD can treat the CFG modification as an additional velocity field during training. The student model learns to match the CFG-guided field directly, internalizing the guidance behavior. As a result, the student model can generate highly aligned images without needing CFG calculations during inference, reducing generation time by half and simplifying deployment.

Summary of Key Benefits

In summary, DanceOPD offers several clear advantages for image generation and editing workflows:

• No Capability Interference: Unifies text-to-image synthesis and image editing in a single model without performance degradation.
• On-Policy Accuracy: Minimizes covariate shift, producing sharper and more coherent images.
• High Computational Efficiency: Reduces teacher query overhead during training using the Semantic-Side Single Query method.
• CFG Integration: Absorbs classifier-free guidance into the student parameters, cutting inference costs.
• Clean Architecture: Replaces multiple specialized networks with a single, compact flow-matching model.