Persistence Forcing: Exploiting Feature Specialization in Pixel-Space Diffusion

Hugging Face Daily Papers Papers
diffusion-transformers pixel-space-diffusion image-generation feature-specialization classifier-free-guidance imagenet

Summary

The paper introduces Persistence Forcing (PerF), which assigns heterogeneous refinement budgets to feature groups in pixel-space Diffusion Transformers, producing persistent global-structure features that guide active detail features. PerF-H reaches FID 1.63 on ImageNet 256x256, outperforming JiT-H with competitive efficiency.

Pixel-space diffusion Transformers (DiTs) directly operate on high-dimensional visual data, yet their hidden representations typically undergo uniform refinement across depth. Natural images, however, are inherently organized at different levels of granularity. Global structure can often be represented compactly, whereas local textures and fine details require richer representations. Motivated by this, we introduce heterogeneous refinement in pixel-space DiTs, assigning different feature groups distinct refinement budgets across depth. Consequently, an ordered feature specialization emerges: sparsely refined features predominantly encode global visual structure, whereas more frequently refined features increasingly specialize toward localized, high-frequency details. We refer to these two groups as persistent and active features, respectively. Building on this emergent specialization, we introduce Persistence Forcing (PerF), which explicitly exploits this persistent--active feature organization for pixel-space image generation. This enables persistent features to continuously condition actively refined features, allowing stable global information to guide the ongoing refinement of finer visual details. During generative sampling, this interaction further induces a meaningful guidance direction that promotes coherent global structure and naturally complements classifier-free guidance. On ImageNet 256times256, PerF-L achieves FID of 1.91, approaching 1.86 of JiT-H with only half the parameters, while PerF-H further achieves FID of 1.63 and 1.76 on ImageNet 256times256 and 512times512, respectively.
Original Article
View Cached Full Text

Cached at: 09/30/26, 12:16 PM

Paper page - Persistence Forcing: Exploiting Feature Specialization in Pixel-Space Diffusion

Source: https://huggingface.co/papers/2609.36014

Abstract

Pixel-spacediffusionTransformers(DiTs)directlyoperateonhigh-dimensionalvisualdata,yettheirhiddenrepresentationstypicallyundergouniformrefinementacrossdepth.Naturalimages,however,areinherentlyorganizedatdifferentlevelsofgranularity.Globalstructurecanoftenberepresentedcompactly,whereaslocaltexturesandfinedetailsrequirericherrepresentations.Motivatedbythis,weintroduceheterogeneousrefinementinpixel-spaceDiTs,assigningdifferentfeaturegroupsdistinctrefinementbudgetsacrossdepth.Consequently,anorderedfeaturespecializationemerges:sparselyrefinedfeaturespredominantlyencodeglobalvisualstructure,whereasmorefrequentlyrefinedfeaturesincreasinglyspecializetowardlocalized,high-frequencydetails.Werefertothesetwogroupsaspersistentandactivefeatures,respectively.Buildingonthisemergentspecialization,weintroducePersistenceForcing(PerF),whichexplicitlyexploitsthispersistent--activefeatureorganizationforpixel-spaceimagegeneration.Thisenablespersistentfeaturestocontinuouslyconditionactivelyrefinedfeatures,allowingstableglobalinformationtoguidetheongoingrefinementoffinervisualdetails.Duringgenerativesampling,thisinteractionfurtherinducesameaningfulguidancedirectionthatpromotescoherentglobalstructureandnaturallycomplementsclassifier-freeguidance.OnImageNet256times256,PerF-LachievesFIDof1.91,approaching1.86ofJiT-Hwithonlyhalftheparameters,whilePerF-HfurtherachievesFIDof1.63and1.76onImageNet256times256and512times512,respectively.

View arXiv pageView PDFProject pageGitHub10Add to collection

Get this paper in your agent:

hf papers read 2609\.36014

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper1

#### ChongWang1024/PerF Updatedabout 10 hours ago • 2

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.36014 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.36014 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion

Hugging Face Daily Papers

PixelDense reframes dense prediction tasks as representation alignment targets for pixel-space diffusion, routing semantic (DINOv2, SAM2) and geometric (DepthAnythingv2, Metric3Dv2) teachers through separate projection streams with an orthogonality penalty. The method improves GenEval, DPG-Bench, and HPSv2.1 scores on PixelGen and DeCo, achieves up to 53.1% PQ gain in partial-noise reconstruction, and trains 1.23x faster than baseline.

Registers Matter for Pixel-Space Diffusion Transformers

Hugging Face Daily Papers

This paper explores the use of register tokens in pixel-space Diffusion Transformers (DiTs), finding they improve feature map quality despite DiTs lacking patch-token outliers. The authors propose Register Guidance, a technique to amplify register contributions for better visual structure.

Adversarial Training for Pixel Diffusion

Hugging Face Daily Papers

This paper presents the first systematic study of adversarial post-training for pixel diffusion models, showing it effectively corrects high-frequency content deficiencies to improve distribution fidelity, coverage, and perceptual quality without altering model architecture.