💻 computer science

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

This paper presents TeleStyle V2, an advanced style transfer model that overcomes the limitations of its predecessor by employing Self-Distillation for versatile content-style reference combinations and Distribution-Matching Distillation to preserve general image editing capabilities, ultimately achieving performance comparable to state-of-the-art commercial models.

Original authors: Shiwen Zhang, Yifan Xu, Haibin Huang, Chi Zhang, Xuelong Li

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shiwen Zhang, Yifan Xu, Haibin Huang, Chi Zhang, Xuelong Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a talented artist who is great at taking a photo of a real person and painting it in the style of Van Gogh. That was the original TeleStyle V1. It was a star performer, but it had a very specific weakness: it only knew how to paint real photos. If you handed it a cartoon drawing and asked it to make it look like a photograph, or if you asked it to paint a real photo in the style of another real photo, it would get confused and fail.

TeleStyle V2 is the artist's big upgrade. The team behind it taught the model to be much more versatile, so it can handle any mix of "real" and "cartoon" inputs. Here is how they did it, explained simply:

1. The "Self-Teaching" Trick (Self-Distillation)

In the old version, the model only practiced on real photos. To fix this, the researchers invented a clever "self-teaching" loop.

Think of the model as a student who has already mastered painting real photos. The researchers told the student: "Okay, take the painting you just made, treat it as a new 'real' photo, and try to paint it in a different style."

By doing this over and over, the model generated its own practice examples where both the starting image and the style were artistic (or both were real). This allowed the model to learn how to handle Stylized-to-Stylized or Real-to-Real tasks, effectively teaching itself to break out of its original box.

2. The "Memory Keeper" (Distribution Matching Distillation)

When you train a model too hard on one specific task (like style transfer), it sometimes starts to forget its other skills, like understanding simple text instructions or keeping the original image looking natural. It's like a chef who learns to make the perfect sushi but forgets how to cook a simple egg.

To stop this "forgetting," the team used a technique called Distribution Matching Distillation (DMD).

The Analogy: Imagine the model is a student taking a new, difficult course. DMD is like a tutor who constantly whispers, "Don't forget the basics you learned in your previous class!"
The Result: This technique ensures the model doesn't lose its ability to follow text commands or keep images looking good. In fact, it made the model so good that it can now do general image editing (like changing a background or adding an object) just as well as the original, untrained version, while still being a master of style transfer.

3. Fixing the "Mix-Up" Problem (Prompt Enhancer)

The researchers noticed a funny glitch: sometimes the model would get the instructions backward. If you showed it a photo of a dog (Content) and a picture of a sunset (Style), it would sometimes ignore the dog and just copy the sunset, or vice versa. It was like a waiter who brings you the wrong menu because they got confused about which dish you ordered.

They tried a complex method called DPO to fix this, but it didn't work. Instead, they found a simple solution: The Prompt Enhancer.

The Analogy: Before the waiter takes the order, they quickly describe the dish to the chef: "The customer wants the dog, not the sunset."
How it works: They use a smart AI assistant (Qwen2.5-VL-7B) to automatically write a short description of the content image and the style image. By adding these descriptions to the instructions, the model immediately understands which image is which, and the mix-up problem disappears.

The Bottom Line

TeleStyle V2 is a major leap forward because:

It can handle any combination of real and artistic images (Real-to-Real, Art-to-Art, Real-to-Art, Art-to-Real).
It didn't lose its general "common sense" for editing images; it actually got better at it.
It fixed the confusion about which image is the "subject" and which is the "style."

The team claims that in terms of keeping the original subject while changing the style, TeleStyle V2 performs just as well as the top-tier commercial model Gemini-3-Pro-Image-Preview (also known as "Nano Banana Pro"), but it is an open-source project. They have also shared their code and project page for others to use.

Technical Summary: TeleStyle V2

Problem Statement

TeleStyle V2 addresses critical limitations observed in its predecessor, TeleStyle V1, specifically regarding the rigidity of input domains and reference handling. While TeleStyle V1 successfully performed content-preserving style transfer, it was trained exclusively on Realistic-and-Stylized (RnS) triplets (photorealistic content references and artistic style references). This training distribution led to three primary failure modes:

Inability to handle stylized content: The model failed when the content reference was an artistic or stylized image.
Inability to handle realistic styles: The model struggled when the style reference was a photorealistic image rather than an artistic one.
Reference Order Confusion: In certain cases, the model ignored the content reference entirely, generating outputs nearly identical to the style reference, despite the use of MS-RoPE and image indices in the input template.

Furthermore, the Supervised Fine-Tuning (SFT) process required for style transfer caused a degradation in the foundation model's general text-guided image editing capabilities and content consistency.

Methodology

1. Self-Distillation for Domain Expansion

To overcome the domain gap, the authors introduced a Self-Distillation data synthesis strategy. Leveraging the existing capabilities of TeleStyle V1, they constructed training triplets covering four distinct reference combinations:

Realistic-and-Realistic (RnR): Using a realistic image as both content and target, with a random realistic image as the style reference.
Stylized-and-Stylized (SnS): Using a single content reference to generate two different stylized targets ( $result_0, result_1$ ) via two different style references. These are then swapped to form new triplets (e.g., $[result_1, style_0, result_0]$ ), creating training data where content, style, and target are all stylized.
Stylized-and-Realistic (SnR): Converting a stylized result back to a realistic target. The authors note this specific task is unstable and recommend using direct text prompts for this conversion rather than the style transfer mode.

This approach allows TeleStyle V2 to generalize across all combinations of realistic and stylized inputs.

2. Distribution Matching Distillation (DMD)

To mitigate content consistency degradation and preserve the foundation model's general editing skills, the authors integrated Distribution Matching Distillation (DMD).

Mechanism: DMD minimizes the reverse Kullback-Leibler (KL) divergence between the real data distribution ( $p_{real}$ ) and the generated data distribution ( $p_{fake}$ ).
Benefits: This technique acts as a regularizer that prevents "catastrophic forgetting" of the base model's capabilities. It enables TeleStyle V2 to perform content-preserving style transfer and general text-guided image editing in a single run. The authors report that DMD reduces inference costs by approximately 10 times while maintaining or enhancing the aesthetic and structural integrity of the output.

3. Prompt Enhancement for Reference Order Confusion

The authors investigated Direct Preference Optimization (DPO) to solve the reference order confusion problem but found it ineffective even with large datasets (400k pairs). Instead, they adopted a Prompt Enhancement strategy:

Implementation: They utilize the Qwen2.5-VL-7B model (already present as the VLM encoder in the Qwen-Image-Edit architecture) to generate a brief textual description of the objects in the content reference.
Effect: This auto-generated "content prompt" is added to the input, providing a semantic hint that guides the model to correctly associate the content reference with the target, effectively resolving the confusion without requiring complex optimization techniques.

Key Contributions

Self-Distillation Mechanism: A novel data synthesis approach that constructs RnR, RnS, SnR, and SnS training triplets, enabling the model to handle mixed modalities (realistic and stylized) for both content and style references.
Integration of DMD: Demonstrating that Distribution Matching Distillation effectively preserves the general text-guided image editing capabilities of the Qwen-Image-Edit series while fixing content consistency issues introduced by SFT.
Prompt Enhancement Solution: Identifying that adding auto-generated content descriptions solves the reference order confusion problem more effectively than DPO.

Results and Evaluation

General Image Editing Capability

Evaluated on GEdit-Bench (using Qwen2.5-VL-72B as the judge), TeleStyleV2-QIE-2509-DMD demonstrated performance on par with, and in some cases superior to, the vanilla Qwen-Image-Edit-2509-DMD across 11 sub-tasks (including background change, color alteration, and subject replacement). Notably, it showed specific improvements in style-changing tasks despite not being explicitly trained on text-guided style transfer data.

Style Transfer Performance

Comparison: TeleStyle V2 was qualitatively compared against the state-of-the-art commercial model gemini-3-pro-image-preview (referred to as "nano banana pro").
Findings: TeleStyle V2 achieved comparable content consistency and, in most cases, superior style similarity to the commercial model.
Quantitative Caution: The authors chose not to release quantitative win-rate statistics against the commercial model, noting that TeleStyle V2's win rate was unexpectedly high, which they suspect may be due to bias in their testing set. They conservatively claim the model performs "on par" with the commercial benchmark.

Significance and Claims

The paper positions TeleStyle V2 as a significant step forward in open-source style transfer, moving beyond the rigid constraints of previous models. By combining self-distillation with distribution matching, the authors claim to have created a model that is not only a specialized style transfer tool but also a robust general text-guided image editor. The work demonstrates that open-source models can achieve performance comparable to top-tier commercial systems (gemini-3-pro-image-preview) in content-preserving style transfer while retaining the flexibility of general image editing. The successful resolution of reference order confusion via simple prompt engineering rather than complex alignment algorithms is highlighted as a practical and effective insight for future model development.