Persistence Forcing

Exploiting Feature Specialization
in Pixel-Space Diffusion

Chong Wang1, Zixuan Fu1, Shiqi Huang1, Siyuan Yang2, Hao Cheng3, Bihan Wen1

1 Nanyang Technological University2 KTH Royal Institute of Technology3 Hebei University of Technology
PaperCodeModel

Some features preserve the big picture.
Others refine the details. Let them work together.

Generated golden puppyGenerated red pandaGenerated giant pandaGenerated tiger
Generated foxGenerated eagleGenerated owlGenerated daisiesGenerated retrieverGenerated lemon and applesGenerated carGenerated coralGenerated mushroomGenerated capitolGenerated clownfishGenerated strawberriesGenerated flowerGenerated pomegranateGenerated butterflyGenerated parrot
Generated directly in pixel space. ImageNet samples from PerF.
PAPER SUMMARY

PerF studies how heterogeneous refinement in pixel-space diffusion Transformers separates persistent features for global structure from active features for fine details. Persistent-to-Active Conditioning and Persistence Guidance use this specialization to improve class-conditional ImageNet generation.

01 / THE DISCOVERY

Different refinement histories.
An ordered feature specialization.

What happens when hidden features are not all updated equally? Varying Transformer width across depth reveals an ordered specialization: sparsely refined features preserve global structure, while frequently refined features contribute finer, higher-frequency details.

HOW HETEROGENEOUS REFINEMENT WORKSVanilla JiT uniformly refines all hidden features, while heterogeneous refinement assigns feature groups different refinement budgets by bypassing selected blocks
Instead of updating every feature in every block, we divide the hidden dimension into groups with different refinement budgets. Low-budget groups bypass more blocks, while high-budget groups remain active throughout the network.
Uniform refinement has no clear specialization; heterogeneous refinement orders contributions from global structure to high-frequency details
Under uniform refinement, feature groups show no comparable ordering. Heterogeneous refinement produces a coarse-to-fine progression.

EXPLORE THE FEATURE CONTRIBUTIONS

From structure to detail

Each image is one feature group's contribution to the final RGB prediction—not an intermediate denoising step.

Refinement budget R

Limited refinement preserves coherent global visual structure.

Feature contribution with refinement budget R = 2
R = 2 · Sparse refinement
02 / THE METHOD

Preserve structure.
Condition active refinement.

Persistence Forcing turns preserved features into useful context. Persistent features bypass a block's transformation and provide spatially aligned, token-wise conditioning for the features that remain active.

Persistent features

Receive fewer updates and retain coherent global visual structure.

Active features

Continue to evolve, using persistent context to refine finer visual details.

Persistence Forcing architecture: persistent-to-active conditioning and predictions with conditioning on and off
Persistent-to-active conditioning also provides a controllable internal signal for sampling-time guidance.

A lightweight interaction. A low-rank projection adds persistent context to the usual timestep and class modulation.

Within a single backbone. The feature roles emerge from heterogeneous refinement rather than being explicitly assigned.

03 / PERSISTENCE GUIDANCE

An internal structural signal.
A complement to CFG.

Compare predictions with and without persistent conditioning. Their difference defines a guidance direction that strengthens the influence of the model's own structural context.

xPG = x∅ + wp (xp − x∅)Persistent conditioning on versus off
No guidanceButterfly sample without guidance
PGButterfly sample with Persistence Guidance
CFGButterfly sample with classifier-free guidance
CFG + PGButterfly sample with combined classifier-free and Persistence Guidance

Same initial noise and class condition. In this example, PG reinforces global structural organization, while CFG strengthens semantic appearance. Their combination preserves both effects.

CFGExternally specified
semantic conditioning

+

PGInternally formed
structural conditioning

04 / THE RESULTS

Better generation,
across scales and resolutions.

PerF improves the corresponding JiT baselines on class-conditional ImageNet generation.

600 training epochs · FID ↓ · IS ↑
JiT and PerF results on ImageNet
ModelParametersFID ↓IS ↑
JiT-B/16131M3.66275.1
PerF-B/16 Ours137M2.81288.3
JiT-L/16459M2.36298.5
PerF-L/16 Ours471M1.91311.2
JiT-H/16953M1.86303.4
PerF-H/16 Ours987M1.63324.5

If you find PerF is useful in your research or applications, please consider giving us a star ⭐ and citing it by the following BibTeX entry.

@misc{wang2026persistenceforcing,
  title={Persistence Forcing: Exploiting Feature Specialization in Pixel-Space Diffusion},
  author={Chong Wang and Zixuan Fu and Shiqi Huang and Siyuan Yang and Hao Cheng and Bihan Wen},
  year={2026},
  eprint={2609.36014},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  doi={10.48550/arXiv.2609.36014},
  url={https://arxiv.org/abs/2609.36014}
}