TriFlowSR: UHD Landmark Super-Resolution

• TriFlowSR is a diffusion-based, reference-guided image super-resolution framework that explicitly aligns low-resolution inputs with high-resolution landmark references using a tri-branch design.
• It integrates Patch-Ref Attention and a Roma-based reference matching strategy to transfer fine-grained textures while mitigating misalignment and generative hallucinations.
• The approach achieves competitive perceptual quality and lower computational costs compared to other diffusion-based RefSR methods on UHD landmark datasets.

Searching arXiv for "TriFlowSR" and directly related super-resolution work. Found relevant results for the exact term and neighboring reference-based super-resolution literature. Searching arXiv for the exact title "Ultra-High-Definition Reference-Based Landmark Image Super-Resolution with Generative Diffusion Prior" and for related benchmark methods named in the data. TriFlowSR is a diffusion-based, reference-guided image super-resolution framework designed specifically for ultra-high-definition landmark images under realistic degradations. It couples a strong generative diffusion prior with explicit feature and pattern matching between a degraded low-resolution input and a high-resolution reference image, and it is trained and evaluated on the Landmark-4K dataset introduced alongside the method (Shi et al., 14 Aug 2025).

1. Problem setting and domain assumptions

TriFlowSR is situated in reference-based image super-resolution (RefSR), where the input consists of a degraded low-resolution image $I^{LR}$ and a high-resolution reference image $I^{Ref}$ of the same or similar scene, and the target is a super-resolved output $I^{SR}$ that remains consistent with the content and structure of $I^{LR}$ while importing semantic and texture information from $I^{Ref}$. In the landmark setting emphasized by the method, the reference is intended to anchor fine-grained structures such as brick patterns, window grids, and sculpture details rather than serve as a generic style exemplar (Shi et al., 14 Aug 2025).

The method is motivated by three constraints that are specific to ultra-high-definition landmark restoration under real-world degradation. First, UHD images at roughly $3$–$4$K impose substantial memory and FLOP costs for diffusion models, so full-resolution processing is impractical and tile-based inference becomes necessary. Second, the degradation model is substantially more complex than bicubic downsampling: the training setup uses Real-ESRGAN’s second-order model with $\times 4$ downsampling, including blur, noise, and JPEG artifacts. Third, prior RefSR datasets such as CUFED5 and WR-SR are comparatively low-resolution, with average resolutions of $418 \times 418$ and $770 \times 770$, so their references often do not contain sufficient fine detail for high-quality UHD restoration.

Within this setting, TriFlowSR is explicitly framed against diffusion-based RefSR pipelines built on ControlNet. The central criticism is that coarse conditioning does not establish precise correspondence between regions in the low-resolution input and regions in the reference image. The concrete failure modes identified for such conditioning include texture misplacement, semantic mismatch, and generative hallucinations when the diffusion prior is insufficiently grounded by aligned reference content.

2. Three-branch architecture and explicit pattern matching

The defining architectural feature of TriFlowSR is its tri-branch organization. The three flows or branches are an SR branch, an LR branch, and a Reference HR branch. The SR branch is a frozen pre-trained text-to-image latent diffusion model built on Stable Diffusion 3 and a rectified flow transformer; it supplies the generative prior and the core latent backbone. The LR branch shares the same architecture type, is pre-trained in stage 1 as an SISR diffusion model using LSDIR and the first $I^{Ref}$0k face images from FFHQ, and is frozen during RefSR training. The Reference HR branch also shares the same architecture type and is the trainable component in stage 2, where it learns to inject aligned reference features into the shared backbone (Shi et al., 14 Aug 2025).

At the system level, the output is written as

$I^{Ref}$1

where $I^{Ref}$2 denotes frozen SR-branch parameters, $I^{Ref}$3 frozen LR-branch parameters, and $I^{Ref}$4 trainable Reference-branch parameters.

The mechanism that distinguishes TriFlowSR from coarse reference conditioning is Patch-Ref Attention. At each attention layer, SR queries attend jointly to SR, LR, and reference tokens: $I^{Ref}$5 This formulation makes the alignment problem explicit at the token level. Each SR query can attend to its own backbone tokens, LR-guided content tokens, and reference-guided texture tokens within the same attention computation. The result is an attention structure in which texture transfer is mediated by learned similarity rather than by simple feature addition.

The reference contribution can also be modulated by scaling the reference keys,

$I^{Ref}$6

so that $I^{Ref}$7 yields behavior close to SISR, while $I^{Ref}$8 corresponds to full reference usage. This control is important because reference quality and alignment vary across datasets and scenes.

3. Reference Matching Strategy and UHD inference

Patch-level attention alone does not resolve the spatial mismatch between a low-resolution input and a reference image acquired from a different viewpoint or scale. TriFlowSR therefore introduces a Reference Matching Strategy that performs pixel-level alignment before tile-based RefSR. The method first bicubic-upscales the LR image to the intended HR size, then downsamples both the upscaled LR image and the reference HR image to a fixed size, given in the supplied description as $I^{Ref}$9. A pre-trained dense matching model, Roma, is then used to estimate correspondences and confidences (Shi et al., 14 Aug 2025): $I^{SR}$0 with

$I^{SR}$1

The mapping $I^{SR}$2 and confidence map $I^{SR}$3 are upscaled to the full target resolution, and the reference HR image is warped with $I^{SR}$4 via $I^{SR}$5: $I^{SR}$6 To suppress unreliable transfers, the warped reference is combined with a mask through the confidence map: $I^{SR}$7 The supplied description states that, in practice, the SISR result is used as the mask rather than a pure white placeholder. High-confidence regions therefore inherit aligned reference texture, while low-confidence regions fall back to SISR content.

For UHD inference, the upscaled LR image and the aligned reference are processed by sliding-window tiling. The reported tile size is $I^{SR}$8 with stride $I^{SR}$9, and c2-blending is used for seamless stitching. This tiled design is not a peripheral implementation detail: it is a necessary condition for applying latent diffusion to $I^{LR}$0–$I^{LR}$1K landmark imagery while preserving local correspondence between LR and reference content.

4. Rectified-flow formulation and training procedure

TriFlowSR uses a rectified-flow variant of diffusion built on Stable Diffusion 3. In the supplied formulation, the noisy path between clean data $I^{LR}$2 and Gaussian noise $I^{LR}$3 is

$I^{LR}$4

and under the rectified-flow choice $I^{LR}$5,

$I^{LR}$6

The model learns a velocity field approximating $I^{LR}$7, and the conditional stage-2 objective is written in latent space as (Shi et al., 14 Aug 2025)

$I^{LR}$8

Training proceeds in two stages. In stage 1, only LR images are used. The LR branch is trained as an SISR diffusion model, with LR images generated by Real-ESRGAN’s second-order degradation model and then bicubic-upsampled to HR size before encoding. The stage-1 datasets are LSDIR and the first $I^{LR}$9k face images from FFHQ, and the training uses mixed resolutions of $I^{Ref}$0 and $I^{Ref}$1.

In stage 2, the SR and LR branches are frozen and only the Reference branch is trained for RefSR. The stage-2 datasets are DL3DV-10K and Inter4K. The optimization settings reported for this stage are AdamW with learning rate $I^{Ref}$2, $I^{Ref}$3k training steps, $I^{Ref}$4 warm-up steps, BF16 precision, and $I^{Ref}$5 NVIDIA H20 GPUs. The backbone uses the VAE encoder and decoder from SD3, and Patch-Ref Attention is inserted into the cross-attention layers where LR and reference latents are fused with the SR backbone features.

5. Landmark-4K dataset and empirical profile

TriFlowSR is introduced together with Landmark-4K, a RefSR dataset for UHD landmark scenarios. The dataset contains $I^{Ref}$6 high-quality landmark images across $I^{Ref}$7 landmark categories, with average resolution about $I^{Ref}$8. For each category, $I^{Ref}$9–$3$0 images are collected from different viewpoints; one is used for self-reference and the others for cross-reference. The acquisition pipeline described in the supplied material begins from about $3$1 images collected from pexels.com, removes images below $3$2K resolution, proportionally downsamples images above $3$3, discards images with noticeable blur or noise, and filters categories to avoid extreme viewpoint differences (Shi et al., 14 Aug 2025).

Dataset	#Images	Avg. Resolution
CUFED5	126	$3$4
WR-SR	80	$3$5
Landmark-4K	185	$3$6

The LR images for Landmark-4K are generated using Real-ESRGAN’s second-order degradation model with $3$7 downsampling. The intended rationale is that UHD landmark references contain the fine-grained structure required for diffusion-based RefSR, whereas low-resolution legacy datasets do not adequately reflect modern smartphone photography.

Quantitatively, the supplied results position TriFlowSR as a method that prioritizes both perceptual quality and competitive structural fidelity. On CUFED5, its reported numbers are PSNR $3$8, SSIM $3$9, LPIPS $4$0, FID $4$1, and DISTS $4$2. On WR-SR, the reported numbers are PSNR $4$3, SSIM $4$4, LPIPS $4$5, FID $4$6, and DISTS $4$7. On Landmark-4K, the reported numbers are PSNR $4$8, SSIM $4$9, LPIPS $\times 4$0, and DISTS $\times 4$1. The comparison described in the source emphasizes that non-diffusion RefSR baselines such as C2-Matching and DATSR may achieve relatively high PSNR on some benchmarks while exhibiting substantially worse LPIPS and DISTS, whereas diffusion-based baselines such as CoSeR and ReFIR are more prone to misaligned detail transfer when reference conditioning is coarse.

The computational profile is also part of the empirical characterization. On $\times 4$2 inputs with an H20 GPU, TriFlowSR is reported at $\times 4$3 TFLOPs and $\times 4$4 s inference time, compared with approximately $\times 4$5 TFLOPs and $\times 4$6 s for SUPIR, approximately $\times 4$7 TFLOPs and $\times 4$8 s for ReFIR, and approximately $\times 4$9 s for CoSeR. In the supplied comparison, it is described as the lowest-TFLOP and fastest among the compared diffusion baselines.

6. Limitations, failure modes, and terminological scope

The limitations stated for TriFlowSR are primarily geometric and data-dependent. The method depends on the availability of a reference that is sufficiently aligned, semantically shared, and high quality. Roma-based matching can fail under extreme viewpoint changes or occlusions, and inaccurate warping can misplace small details such as faces, fingers, or ornaments. On WR-SR, where references are often obtained by Google search and may be poorly aligned, TriFlowSR’s PSNR and SSIM are reported as slightly lower than some bicubic-based CNN RefSR methods, even though its perceptual metrics remain better. A further failure mode is degeneration toward SISR when $418 \times 418$0 is set very low or when the Reference branch is too shallow (Shi et al., 14 Aug 2025).

Within the supplied literature, the name “TriFlowSR” also appears in other, explicitly distinct senses. One supplied explanation treats TriFlowSR as a tri-channel $418 \times 418$1 flow super-resolution framework aligned with “Vision-Informed Flow Image Super-Resolution with Quaternion Spatial Modeling and Dynamic Flow Convolution,” where quaternion spatial modeling represents orthogonal velocity components and dynamic flow convolution models droplet-like flow morphology (Cao et al., 2024). Another supplied explanation treats “TriFlowSR” as a hypothetical triad-aware method inspired by Triadic Orthogonal Decomposition, emphasizing donor, recipient, and catalyst roles in spectral energy transfer across frequencies (Yeung et al., 2024). A further supplied note states that “TriFlowSR” is not a name that appears in the TriTPP paper, although an interpretive connection was drawn there to triangular flow models for fast sampling and reconstruction in temporal point processes (Shchur et al., 2020).

In the formal arXiv record supplied here, however, TriFlowSR is the title of the UHD landmark RefSR framework based on a frozen Stable Diffusion 3 prior, an LR branch, a Reference HR branch, Patch-Ref Attention, and a Roma-based Reference Matching Strategy (Shi et al., 14 Aug 2025).