DualFit: Two-Stage Virtual Try-On

• The paper introduces DualFit, a two-stage framework combining explicit garment-preserving warping with fidelity-preserving synthesis to maintain logos and intricate patterns.
• In the first stage, a learned flow field aligns garment parts using local and global parsing, while the second stage employs a Res-UNet with seam-aware inpainting to repair discontinuities.
• Empirical results on VITON-HD demonstrate state-of-the-art performance with improved metrics like SSIM, FID, and LPIPS, balancing detail fidelity and perceptual realism.

Searching arXiv for the specified paper and closely related VTON work to ground the article with citations. DualFit is a virtual try-on (VTON) method introduced in "DualFit: A Two-Stage Virtual Try-On via Warping and Synthesis" (Tran et al., 16 Aug 2025). It is designed to address a specific failure mode of recent diffusion-based warping-free systems: the loss of high-frequency garment content such as logos, printed text, and intricate patterns. The method combines an explicit garment-preserving warping stage with a fidelity-preserving synthesis stage. In the first stage, the target garment is aligned to the person image with a learned flow field; in the second, a synthesis module blends the warped garment with preserved human regions while regenerating only where necessary, particularly around garment seams (Tran et al., 16 Aug 2025). The resulting pipeline is positioned between artifact-prone classical warping-based VTON and detail-attenuating warping-free diffusion methods.

1. Problem setting and design objective

DualFit is motivated by the observation that recent warping-free methods built on latent diffusion improve perceptual quality but often wash out fine-grained garment details. The abstract and technical description identify two causes: information bottlenecks in VAEs and denoising U-Nets operating in latent space, where fine details are attenuated, and stochastic sampling, which can further blur determinism (Tran et al., 16 Aug 2025). In practice, this matters because logos and printed text are described as critical for brand integrity and customer trust.

The method therefore targets a narrow but consequential objective: preserving the actual appearance of the in-shop garment while still producing visually seamless try-on results. Its remedy is explicitly hybrid. Stage 1 transforms the actual pixels of the in-shop garment, maintaining textures; Stage 2 uses preserved human regions and a seam-focused inpainting mask to blend the warped garment seamlessly, avoiding the unnatural boundaries associated with classical warp-and-render pipelines (Tran et al., 16 Aug 2025).

This framing places DualFit in a trade-off space already familiar in VTON research. Warping preserves texture but can introduce hard boundaries and misalignment artifacts; warping-free synthesis improves smoothness but can erase brand-critical graphics. DualFit is presented as a two-stage compromise that seeks both reconstruction fidelity and perceptual realism (Tran et al., 16 Aug 2025).

2. Two-stage pipeline

The pipeline has two stages: warping and fidelity-preserving synthesis (Tran et al., 16 Aug 2025).

In Stage 1, a learned appearance flow aligns the flat garment $G$ to the person image $I$ using pose and densepose cues. Local flows are predicted for garment subregions—left sleeve, torso, and right sleeve—and a global garment parsing map assembles these parts into a final warped garment $G'$ (Tran et al., 16 Aug 2025). The role of local-part decomposition is to reduce deformation artifacts.

In Stage 2, DualFit constructs a preserved-region input $I'$ by removing the upper body, excluding head and hair, and overlaying the warped garment $G'$. It also defines an inpainting mask $M$ focused on inter-part seams and key body parts, specifically hands and neck. A Res-UNet then synthesizes the final try-on output $O$, conditioned on $I'$, $G'$, the flat garment $G$, pose $I$0, and $I$1 (Tran et al., 16 Aug 2025). According to the method description, this generator preserves aligned details from $I$2 while regenerating only the masked regions for continuity and realism.

A concise statement of the system is therefore:

• Stage 1: preserve garment appearance by geometric alignment of actual garment pixels.
• Stage 2: repair seam discontinuities and body-part interactions with targeted synthesis.

This suggests that DualFit treats warping not as a complete rendering solution but as a high-fidelity prior that the synthesis network edits only where warping is structurally weak.

3. Warping stage: learned flow, local-part composition, and GP-VTON lineage

The warping stage takes as inputs the person image $I$3, in-shop garment $I$4, human densepose $I$5, and pose heatmap $I$6 (Tran et al., 16 Aug 2025). The learned flow field is defined as a dense displacement field $I$7 over the pixel domain $I$8 mapping garment coordinates to person coordinates. Warping is implemented with a spatial transformer using bilinear sampling:

$I$9

and, more explicitly,

$G'$0

where $G'$1 are the 4 bilinear neighbors and $G'$2 are the standard bilinear weights (Tran et al., 16 Aug 2025).

Rather than learning a single monolithic flow, DualFit predicts local flows $G'$3, $G'$4, and $G'$5 for the left sleeve, torso, and right sleeve, together with a global garment parsing map $G'$6. The final warped garment is assembled as

$G'$7

The architecture is inherited from the Local-Flow Global-Parsing (LFGP) module of GP-VTON. It uses Feature Pyramid Networks (FPNs) to extract multi-scale features from a person stream conditioned by $G'$8 and $G'$9, and from a garment stream with garment parsing; cascaded LFGP blocks refine local flows at multiple scales; and a global garment parsing head assigns pixels to parts and resolves overlaps during assembly (Tran et al., 16 Aug 2025). Conditioning signals include pose heatmaps, densepose maps, and garment parsing.

For training robustness, DualFit adopts Dynamic Gradient Truncation (DGT) from GP-VTON. The paper states that DGT adapts gradient masking based on wearing style, such as tucked-in versus tucked-out, mitigating texture squeezing and stretching by selectively truncating or propagating gradients in preserved regions (Tran et al., 16 Aug 2025). Warping supervision also follows GP-VTON. While the DualFit paper emphasizes the synthesis stage, it notes that warping typically combines part-wise correspondence losses with flow regularization, and gives a canonical smoothness term:

$I'$0

This stage situates DualFit within a GP-VTON-derived geometric alignment lineage rather than a purely generative one.

4. Fidelity-preserving synthesis and seam-aware inpainting

The second stage is the central innovation of DualFit. Its purpose is not to redraw the garment wholesale, but to preserve aligned garment details from $I'$1 and regenerate only those regions where warping alone produces discontinuities or implausible boundaries (Tran et al., 16 Aug 2025).

The preserved-region input is constructed using predicted upper-body parsing $I'$2. The upper body, except head and hair, is removed from the person image $I'$3, and the warped garment $I'$4 is overlaid to form $I'$5 (Tran et al., 16 Aug 2025). This creates an input in which large regions are already fixed by direct garment placement rather than by synthesis.

To determine where synthesis should occur, DualFit computes narrow boundary bands for left sleeve $I'$6, right sleeve $I'$7, and torso $I'$8:

$I'$9

and

$G'$0

where $G'$1 is a $G'$2 square kernel and $G'$3 controls band thickness; the reported choice is $G'$4 (Tran et al., 16 Aug 2025). The final inpainting mask is then

$G'$5

where $G'$6, $G'$7, and $G'$8 denote left hand, right hand, and neck masks (Tran et al., 16 Aug 2025).

The final output is generated by a Res-UNet:

$G'$9

A stated reason for conditioning on the flat garment $M$0 is that it provides high-frequency appearance cues that help the network faithfully reconstruct logos and text (Tran et al., 16 Aug 2025). For intuition, the paper gives the classical compositing equation

$M$1

but clarifies that DualFit learns a superior blending through $M$2, preserving $M$3 in unmasked regions and regenerating only near seams and necessary body parts (Tran et al., 16 Aug 2025).

The network is described as a U-Net with residual blocks and skip connections for spatial detail retention, with batch normalization and optional dropout for stable training and generalization, and it operates directly at full resolution, $M$4 (Tran et al., 16 Aug 2025). This full-resolution processing is important because the paper explicitly contrasts DualFit with latent-space approaches in which high-frequency detail is attenuated.

5. Training protocol, data, and empirical results

The generator objective combines pixel, perceptual, and adversarial terms:

$M$5

with

$M$6

and

$M$7

The paper states that $M$8 is key for copying aligned details from $M$9, while $O$0 provides standard discriminator supervision for realism (Tran et al., 16 Aug 2025). During training, the warped garment is spatially aligned with the ground-truth try-on image $O$1 so that the model learns to preserve garment fidelity by directly copying high-frequency content from $O$2 in preserved regions and synthesizing masked regions smoothly.

The dataset is VITON-HD, with 11,647 training and 2,032 test pairs of upper-body garments at $O$3 resolution for both person and garment images (Tran et al., 16 Aug 2025). Preprocessing includes pose estimation, densepose extraction, upper-body parsing, and garment-part parsing. The warping module predicts flows at $O$4, upsamples flows to $O$5, and then warps the original-resolution garment; the try-on module runs at full $O$6 (Tran et al., 16 Aug 2025).

The reported compute setup is:

Component	Hardware / schedule
Warping module	2× RTX A6000, 150 epochs, lr=$O$7, batch=2/GPU
Try-on module	4× RTX A6000, 200 epochs, lr=$O$8, batch=4/GPU

Evaluation uses PSNR and SSIM for reconstruction, FID for perceptual realism, and LPIPS and DISTS as hybrid metrics (Tran et al., 16 Aug 2025). On VITON-HD, DualFit reports the following quantitative performance:

Metric	Value
PSNR	24.856
SSIM	0.913
FID	5.261
LPIPS	0.064
DISTS	0.046

The paper states that these results are state-of-the-art across categories and that DualFit outperforms warping-free methods such as CatVTON and IDM-VTON in fidelity and realism, while surpassing the warping-based GP-VTON in reconstruction and avoiding its boundary artifacts (Tran et al., 16 Aug 2025). Qualitatively, DualFit is reported to preserve logos, text, and fine patterns while producing smooth seams and realistic draping.

Ablation results isolate two key components. Conditioning on the flat garment $O$9 improves SSIM from $I'$0 to $I'$1, FID from $I'$2 to $I'$3, and LPIPS from $I'$4 to $I'$5 (Tran et al., 16 Aug 2025). For the seam band thickness parameter, $I'$6 yields the best trade-off, with SSIM $I'$7, FID $I'$8, and LPIPS $I'$9; the paper states that too thin or too thick bands degrade seam quality (Tran et al., 16 Aug 2025).

The runtime discussion is qualitative rather than fully benchmarked in the provided material: the model is described as significantly lighter than latent diffusion-based baselines, and its reported FPS is said to indicate higher practicality compared to heavy warping-free U-Nets (Tran et al., 16 Aug 2025).

6. Relation to prior VTON paradigms, limitations, and scope

DualFit is explicitly described as a complement to warping-free diffusion rather than a rejection of synthesis-based VTON. Its argument is that latent-space detail loss can be sidestepped by preserving true garment pixels with learned flow, while a lightweight generator repairs only what warping cannot, including seams and occlusions (Tran et al., 16 Aug 2025). This distinguishes it from pure warping-based systems, where naïve rendering can produce hard boundaries, and from pure warping-free systems, where stochastic latent generation can blur garment-specific details.

The paper also makes the design trade-off explicit: introducing warping adds flow estimation complexity, but this cost is repaid in detail fidelity; synthesis is kept lightweight and targeted, yielding better speed than diffusion and better realism than pure warping (Tran et al., 16 Aug 2025). This suggests a broader methodological principle in VTON: direct pixel transport and localized generation need not be competing paradigms if the generation region is tightly constrained.

The method has several stated limitations. It depends on parsing and segmentation robustness, and errors in upper-body parsing or garment-part segmentation can propagate to mask construction and warping assembly (Tran et al., 16 Aug 2025). Challenging cases include extreme poses, occlusions, complex backgrounds, and highly nonrigid fabrics, all of which can stress flow estimation and parsing and lead to local distortions or incomplete coverage. Very long garments and cross-category try-on are beyond the current upper-body focus (Tran et al., 16 Aug 2025).

Input requirements also constrain deployment. The method requires a person image $G'$0, flat garment $G'$1, pose heatmap $G'$2, densepose $G'$3, and human parsing for masks (Tran et al., 16 Aug 2025). User control is available through pose and densepose conditioning, and mask thickness $G'$4 controls the extent of seam regeneration. The project page is listed at https://uark-aicv.github.io/DualFit (Tran et al., 16 Aug 2025).

A common misconception would be to treat DualFit as merely another warping-based VTON system. The technical description does not support that simplification. Its synthesis stage is not a generic renderer but a fidelity-preserving module organized around preserved-region guidance and seam-aware inpainting. Conversely, it would also be inaccurate to classify it as warping-free, since the method relies fundamentally on learned flow and part-wise spatial transformation (Tran et al., 16 Aug 2025). Its contribution lies in coupling the two regimes so that each addresses the failure modes of the other.