TeleStyle V2: Beyond Content-Preserving Style Transfer with Self-Distillation and Distribution-Matching-Distillation

Chi Zhang; Haibin Huang; Shiwen Zhang; Xuelong Li; Yifan Xu

arxiv: 2606.20709 · v1 · pith:VKSSVDNXnew · submitted 2026-06-16 · 💻 cs.CV

TeleStyle V2: Beyond Content-Preserving Style Transfer with Self-Distillation and Distribution-Matching-Distillation

Shiwen Zhang , Yifan Xu , Haibin Huang , Chi Zhang , Xuelong Li This is my paper

Pith reviewed 2026-06-27 01:13 UTC · model grok-4.3

classification 💻 cs.CV

keywords style transferself-distillationdistribution matching distillationimage editingcontent consistencyprompt enhancerreference order confusion

0 comments

The pith

TeleStyle V2 uses self-distillation on V1 outputs to support style transfer for every combination of realistic and stylized content-style references while retaining the base model's text-guided editing performance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper extends TeleStyle V1, which handled only photorealistic content with artistic style references, by creating self-distilled training triplets that cover all four reference-type pairs: realistic-realistic, realistic-stylized, stylized-realistic, and stylized-stylized. Distribution Matching Distillation is then applied after supervised fine-tuning to counteract content consistency losses and keep the model's general text-guided image editing ability intact. The resulting TeleStyle V2 model matches or exceeds the base Qwen-Image-Edit-2509-DMD on editing benchmarks and reaches performance levels comparable to commercial systems such as gemini-3-pro-image-preview. A prompt enhancer is added to correct reference-order confusion that appeared in the earlier version. TeleStyle V2 relies on the Qwen2.5-VL-7B VLM encoder to generate free content and style prompts.

Core claim

Trained on self-distilled triplets that span all four content-style reference combinations and refined via Distribution Matching Distillation, TeleStyle V2 produces content-consistent and style-consistent outputs for realistic-realistic, realistic-stylized, stylized-realistic, and stylized-stylized pairs while preserving the foundation model's general text-guided image editing capability at a level on par with Qwen-Image-Edit-2509-DMD.

What carries the argument

Self-Distillation data synthesis that generates training triplets from TeleStyle V1 outputs, paired with Distribution Matching Distillation to restore editing performance after supervised fine-tuning.

If this is right

The model now handles artistic content paired with realistic style references without the failures seen in V1.
Distribution Matching Distillation prevents the content consistency drop that normally follows supervised fine-tuning on style data.
TeleStyle V2 achieves editing performance at least equal to Qwen-Image-Edit-2509-DMD across standard text-guided tasks.
A prompt enhancer corrects the reference order confusion observed in TeleStyle V1.
Style transfer quality reaches levels comparable to the commercial gemini-3-pro-image-preview system.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same self-distillation loop could be reused to extend other asymmetric editing models to additional reference domains.
Distribution Matching Distillation might serve as a lightweight alternative to full reinforcement learning when preserving base capabilities after targeted fine-tuning.
Using the VLM encoder for free prompt generation suggests a route to reduce reliance on manual prompt engineering in multi-reference editing pipelines.

Load-bearing premise

The triplets self-distilled from TeleStyle V1 are high-quality and diverse enough to let the model generalize across all reference-type combinations without adding new inconsistencies or degrading base editing ability.

What would settle it

Quantitative evaluation on stylized-content with realistic-style pairs showing either visible content inconsistency or lower general editing scores than the unmodified Qwen-Image-Edit-2509-DMD baseline.

read the original abstract

Given a content reference and a style reference, content-preserving style transfer requires the model to generate stylized outputs with content and style consistency. We introduced TeleStyle V1 to tackle this problem. However, TeleStyle V1 is trained with photorealistic content reference and artistic style reference, which makes it incapable to cope with artistic content reference and realistic style reference in most cases. In this paper, we designed a Self-Distillation data synthesis strategy to construct such triplets from TeleStyle V1. Trained with such self-distilled triplets, our TeleStyle V2 supports Content-Style references in the forms of Realistic-and-Realistic (RnR), Realistic-and-Stylized (RnS), Stylized-and-Realistic (SnR), Stylized-and-Stylized (SnS). In addition, we found Distribution Matching Distillation could preserve the general text-guided image editing capability of the foundation model and fix the content consistency degradation caused by SFT process. Through quantitative evaluations, our TeleStyleV2-QIE-2509-DMD performs at least on par with Qwen-Image-Edit-2509-DMD, demonstrating strong general image editing skills beyond content-preserving style transfer. We observed the content/style reference order confusion problem in TeleStyle V1 and further introduced prompt enhancer to solve it. TeleStyle V2 uses Qwen-Image-Edit's VLM encoder, Qwen2.5-VL-7B, to generate content prompt and style prompt for free. TeleStyle V2 could achieve comparable style transfer performance with state-of-the-art commercial model, gemini-3-pro-image-preview.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

V2 mainly adds self-distillation from V1 to cover SnR and SnS cases plus DMD to hold onto base editing ability, but the data quality assumption looks shaky from the abstract.

read the letter

The core update is straightforward: V1 only handled realistic content with artistic style, so they use V1 to generate new triplets for the other three combinations and train V2 on them. They also add distribution-matching distillation to keep the underlying model's text-guided editing intact and a prompt enhancer to fix reference order mix-ups. The VLM encoder from Qwen supplies the prompts at no extra cost.

This does expand the supported reference types to all four combinations and claims on-par results with Qwen-Image-Edit and Gemini-3-pro. The DMD step is a reasonable way to mitigate the usual SFT degradation.

The soft spot is the self-distillation step itself. V1 was already weak on artistic content, so feeding its outputs back in for SnR and SnS risks carrying forward the same inconsistencies or artifacts. The abstract gives no numbers on triplet quality, diversity, or consistency checks, and the performance claims rest on that unverified data. Without those details the generalization argument stays provisional.

The work is aimed at people building practical image-editing pipelines who need flexible style transfer. It is incremental engineering rather than a new direction, but the pipeline is concrete enough that a referee could check the ablations and metrics in the full paper. I would send it to review.

Referee Report

2 major / 2 minor

Summary. The paper introduces TeleStyle V2 as an extension of TeleStyle V1 for content-preserving style transfer. It uses a Self-Distillation strategy to synthesize training triplets from V1 outputs, enabling the model to handle all four reference-type combinations (RnR, RnS, SnR, SnS). Distribution-Matching-Distillation (DMD) is applied to preserve general text-guided image editing capabilities of the base model (Qwen-Image-Edit) and mitigate content consistency degradation from supervised fine-tuning. A prompt enhancer addresses reference order confusion, and Qwen2.5-VL-7B is used to generate content and style prompts. The authors claim that TeleStyleV2-QIE-2509-DMD performs at least on par with Qwen-Image-Edit-2509-DMD in quantitative evaluations and achieves comparable style transfer performance to the commercial model gemini-3-pro-image-preview.

Significance. If the self-distilled data quality and DMD preservation claims hold, the work would offer a practical route to generalize style transfer beyond the RnS setting of V1 while retaining broad editing skills. The combination of self-distillation for data expansion and DMD for capability retention is a potentially useful technical contribution in the image editing literature. However, the significance is limited by the absence of independent validation metrics on the generated triplets and by the circular dependence on V1 outputs.

major comments (2)

[Self-Distillation data synthesis strategy] The central generalization claim (support for SnR and SnS) rests on the Self-Distillation data synthesis strategy producing high-quality, diverse triplets without propagating V1 artifacts. No quantitative metrics on triplet quality, consistency scores, or diversity for the SnR/SnS cases are reported, leaving the weakest assumption untested.
[Quantitative evaluations] The quantitative evaluation claim that TeleStyleV2-QIE-2509-DMD performs on par with Qwen-Image-Edit-2509-DMD is load-bearing for the broader editing capability assertion, yet the abstract (and provided text) supplies no specific metrics, baselines, datasets, or ablation results to support it.

minor comments (2)

[Abstract] The abstract states that quantitative evaluations were performed but provides none of the actual numbers, tables, or dataset details; these should be summarized with key numbers even in the abstract.
[Introduction] Notation for the four reference combinations (RnR, RnS, SnR, SnS) is introduced without an explicit definition table or diagram, which would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive feedback. We address the major comments point by point below and commit to revisions that directly strengthen the manuscript's claims regarding data quality and evaluation transparency.

read point-by-point responses

Referee: [Self-Distillation data synthesis strategy] The central generalization claim (support for SnR and SnS) rests on the Self-Distillation data synthesis strategy producing high-quality, diverse triplets without propagating V1 artifacts. No quantitative metrics on triplet quality, consistency scores, or diversity for the SnR/SnS cases are reported, leaving the weakest assumption untested.

Authors: We agree that explicit quantitative validation of the self-distilled triplets for SnR and SnS would provide stronger support for the generalization claim and reduce reliance on downstream results alone. Although the V1 outputs were previously validated and the final model performance on SnR/SnS tasks offers indirect evidence, we will add a new analysis subsection reporting consistency scores (content and style) and diversity metrics computed via Qwen2.5-VL-7B on sampled SnR/SnS triplets, along with a brief discussion of artifact propagation checks. revision: yes
Referee: [Quantitative evaluations] The quantitative evaluation claim that TeleStyleV2-QIE-2509-DMD performs on par with Qwen-Image-Edit-2509-DMD is load-bearing for the broader editing capability assertion, yet the abstract (and provided text) supplies no specific metrics, baselines, datasets, or ablation results to support it.

Authors: The full manuscript contains quantitative comparisons on image editing benchmarks demonstrating that TeleStyleV2-QIE-2509-DMD performs at least on par with Qwen-Image-Edit-2509-DMD. However, these details are not sufficiently highlighted in the abstract or early sections. We will revise the abstract to include key metrics and add an expanded evaluation section with explicit numbers, baselines, datasets, and ablation studies to make the claim self-contained and verifiable. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper's training pipeline uses self-distillation from TeleStyle V1 to synthesize triplets for new reference combinations, followed by DMD to retain base editing capabilities from the foundation model. Central claims rest on quantitative evaluations showing TeleStyleV2-QIE-2509-DMD performs on par with external models (Qwen-Image-Edit-2509-DMD, gemini-3-pro-image-preview). No equations, fitted parameters, or uniqueness theorems reduce the reported results to the inputs by construction. Self-generated data serves as a training augmentation rather than a definitional or statistical equivalence, and external benchmarks provide independent verification. The derivation chain remains self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are stated in the provided text.

pith-pipeline@v0.9.1-grok · 5849 in / 999 out tokens · 31786 ms · 2026-06-27T01:13:34.864957+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 7 linked inside Pith

[1]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923,

Pith/arXiv arXiv
[2]

Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

Pith/arXiv arXiv
[3]

Gpt-4o system card.arXiv preprint arXiv:2410.21276,

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

Pith/arXiv arXiv
[4]

Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761,

Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761,

Pith/arXiv arXiv
[5]

Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

Pith/arXiv arXiv
[6]

Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M¨ uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

Pith/arXiv arXiv
[7]

Omniconsistency: Learning style-agnostic consistency from paired stylization data.arXiv preprint arXiv:2505.18445,

Yiren Song, Cheng Liu, and Mike Zheng Shou. Omniconsistency: Learning style-agnostic consistency from paired stylization data.arXiv preprint arXiv:2505.18445,

arXiv
[8]

Qwen-image technical report.arXiv preprint arXiv:2508.02324,

Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324,

Pith/arXiv arXiv
[9]

Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning.arXiv preprint arXiv:2203.05928,

Shiwen Zhang. Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning.arXiv preprint arXiv:2203.05928,

arXiv
[10]

V4d: 4d convolutional neural networks for video-level representation learning

Shiwen Zhang, Sheng Guo, Weilin Huang, Matthew R Scott, and Limin Wang. V4d: 4d convolutional neural networks for video-level representation learning. InInternational Conference on Learning Representations, 2020a. Shiwen Zhang, Sheng Guo, Limin Wang, Weilin Huang, and Matthew Scott. Knowledge integration networks for action recognition. InProceedings of t...

arXiv
[11]

Cdst: Color disentangled style transfer for universal style reference customization.arXiv preprint arXiv:2506.13770,

Shiwen Zhang, Zhuowei Chen, Lang Chen, and Yanze Wu. Cdst: Color disentangled style transfer for universal style reference customization.arXiv preprint arXiv:2506.13770,

arXiv
[12]

Qwenstyle: Content-preserving style transfer with qwen-image-edit.arXiv preprint arXiv:2601.06202, 2026a

Shiwen Zhang, Haibin Huang, Chi Zhang, and Xuelong Li. Qwenstyle: Content-preserving style transfer with qwen-image-edit.arXiv preprint arXiv:2601.06202, 2026a. Shiwen Zhang, Haoyuan Wang, Xianghao Zang, Haibin Huang, Chi Zhang, and Xuelong Li. Style-ccl: Content-preserving style transfer via curriculum continual learning.arXiv preprint arXiv:2606, 2026b....

arXiv

[1] [1]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923,

Pith/arXiv arXiv

[2] [2]

Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685,

Pith/arXiv arXiv

[3] [3]

Gpt-4o system card.arXiv preprint arXiv:2410.21276,

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

Pith/arXiv arXiv

[4] [4]

Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761,

Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761,

Pith/arXiv arXiv

[5] [5]

Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

Pith/arXiv arXiv

[6] [6]

Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M¨ uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

Pith/arXiv arXiv

[7] [7]

Omniconsistency: Learning style-agnostic consistency from paired stylization data.arXiv preprint arXiv:2505.18445,

Yiren Song, Cheng Liu, and Mike Zheng Shou. Omniconsistency: Learning style-agnostic consistency from paired stylization data.arXiv preprint arXiv:2505.18445,

arXiv

[8] [8]

Qwen-image technical report.arXiv preprint arXiv:2508.02324,

Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324,

Pith/arXiv arXiv

[9] [9]

Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning.arXiv preprint arXiv:2203.05928,

Shiwen Zhang. Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning.arXiv preprint arXiv:2203.05928,

arXiv

[10] [10]

V4d: 4d convolutional neural networks for video-level representation learning

Shiwen Zhang, Sheng Guo, Weilin Huang, Matthew R Scott, and Limin Wang. V4d: 4d convolutional neural networks for video-level representation learning. InInternational Conference on Learning Representations, 2020a. Shiwen Zhang, Sheng Guo, Limin Wang, Weilin Huang, and Matthew Scott. Knowledge integration networks for action recognition. InProceedings of t...

arXiv

[11] [11]

Cdst: Color disentangled style transfer for universal style reference customization.arXiv preprint arXiv:2506.13770,

Shiwen Zhang, Zhuowei Chen, Lang Chen, and Yanze Wu. Cdst: Color disentangled style transfer for universal style reference customization.arXiv preprint arXiv:2506.13770,

arXiv

[12] [12]

Qwenstyle: Content-preserving style transfer with qwen-image-edit.arXiv preprint arXiv:2601.06202, 2026a

Shiwen Zhang, Haibin Huang, Chi Zhang, and Xuelong Li. Qwenstyle: Content-preserving style transfer with qwen-image-edit.arXiv preprint arXiv:2601.06202, 2026a. Shiwen Zhang, Haoyuan Wang, Xianghao Zang, Haibin Huang, Chi Zhang, and Xuelong Li. Style-ccl: Content-preserving style transfer via curriculum continual learning.arXiv preprint arXiv:2606, 2026b....

arXiv