Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fast, faithful and photorealistic diffusion-based image super-resolution with enhanced Flow Map models

Maxence Noble, Gonzalo Iñaki Quintana, Benjamin Aubin, Clément Chadebec

TL;DR

FlowMapSR is proposed, a novel diffusion-based framework for image super-resolution explicitly designed for efficient inference that achieves a better balance between reconstruction faithfulness and photorealism than recent state-of-the-art methods for both x4 and x8 upscaling, while maintaining competitive inference time.

Abstract

Diffusion-based image super-resolution (SR) has recently attracted significant attention by leveraging the expressive power of large pre-trained text-to-image diffusion models (DMs). A central practical challenge is resolving the trade-off between reconstruction faithfulness and photorealism. To address inference efficiency, many recent works have explored knowledge distillation strategies specifically tailored to SR, enabling one-step diffusion-based approaches. However, these teacher-student formulations are inherently constrained by information compression, which can degrade perceptual cues such as lifelike textures and depth of field, even with high overall perceptual quality. In parallel, self-distillation DMs, known as Flow Map models, have emerged as a promising alternative for image generation tasks, enabling fast inference while preserving the expressivity and training stability of standard DMs. Building on these developments, we propose FlowMapSR, a novel diffusion-based framework for image super-resolution explicitly designed for efficient inference. Beyond adapting Flow Map models to SR, we introduce two complementary enhancements: (i) positive-negative prompting guidance, based on a generalization of classifier free-guidance paradigm to Flow Map models, and (ii) adversarial fine-tuning using Low-Rank Adaptation (LoRA). Among the considered Flow Map formulations (Eulerian, Lagrangian, and Shortcut), we find that the Shortcut variant consistently achieves the best performance when combined with these enhancements. Extensive experiments show that FlowMapSR achieves a better balance between reconstruction faithfulness and photorealism than recent state-of-the-art methods for both x4 and x8 upscaling, while maintaining competitive inference time. Notably, a single model is used for both upscaling factors, without any scale-specific conditioning or degradation-guided mechanisms.

Fast, faithful and photorealistic diffusion-based image super-resolution with enhanced Flow Map models

TL;DR

FlowMapSR is proposed, a novel diffusion-based framework for image super-resolution explicitly designed for efficient inference that achieves a better balance between reconstruction faithfulness and photorealism than recent state-of-the-art methods for both x4 and x8 upscaling, while maintaining competitive inference time.

Abstract

Diffusion-based image super-resolution (SR) has recently attracted significant attention by leveraging the expressive power of large pre-trained text-to-image diffusion models (DMs). A central practical challenge is resolving the trade-off between reconstruction faithfulness and photorealism. To address inference efficiency, many recent works have explored knowledge distillation strategies specifically tailored to SR, enabling one-step diffusion-based approaches. However, these teacher-student formulations are inherently constrained by information compression, which can degrade perceptual cues such as lifelike textures and depth of field, even with high overall perceptual quality. In parallel, self-distillation DMs, known as Flow Map models, have emerged as a promising alternative for image generation tasks, enabling fast inference while preserving the expressivity and training stability of standard DMs. Building on these developments, we propose FlowMapSR, a novel diffusion-based framework for image super-resolution explicitly designed for efficient inference. Beyond adapting Flow Map models to SR, we introduce two complementary enhancements: (i) positive-negative prompting guidance, based on a generalization of classifier free-guidance paradigm to Flow Map models, and (ii) adversarial fine-tuning using Low-Rank Adaptation (LoRA). Among the considered Flow Map formulations (Eulerian, Lagrangian, and Shortcut), we find that the Shortcut variant consistently achieves the best performance when combined with these enhancements. Extensive experiments show that FlowMapSR achieves a better balance between reconstruction faithfulness and photorealism than recent state-of-the-art methods for both x4 and x8 upscaling, while maintaining competitive inference time. Notably, a single model is used for both upscaling factors, without any scale-specific conditioning or degradation-guided mechanisms.
Paper Structure (55 sections, 26 equations, 18 figures, 6 tables)

This paper contains 55 sections, 26 equations, 18 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Preliminaries on Flow Map models
  4. Background on Flow Matching
  5. Interpolant & ODE.
  6. Training & Inference.
  7. FM enhancing via CFG.
  8. Background on Flow Map models
  9. Definition & parameterization.
  10. Self-distillation losses.
  11. Training & Inference.
  12. Generalization of CFG to the Flow Map setting
  13. Flow matching loss.
  14. Self-distillation losses.
  15. Training & Inference.
  16. ...and 40 more sections

Figures (18)

  • Figure 1: Qualitative super-resolution results of the Shortcut variant of FlowMapSR for $\times 4$ and $\times 8$ upscaling. A single Flow Map model is used for both upscaling factors, with 2 inference steps.
  • Figure 2: Qualitative comparison of positive-negative CFG enhancement applied to LSD-, ESD-, and SSD-based FlowMapSR for $\times 4$ upscaling. The HR reference image exhibits sharp edges and rich high-frequency details. The base FlowMapSR model ("Without CFG") successfully recovers structural details but lacks perceptual sharpness. Applying positive–negative CFG improves visual realism in the Shortcut formulation (e.g., more natural hair textures), while it introduces noticeable degradations and artefacts in the Lagrangian and Eulerian formulations.
  • Figure 3: Qualitative comparison between FlowMapSR and competing SR methods for $\times 4$ (rows 1–4) and $\times 8$ (rows 5–8) upscaling. Rows 1–3 and 5–7 show examples from DIV2K-Val, while rows 4 and 8 correspond to RealSR samples.
  • Figure 4: Qualitative comparison between FlowMapSR and competing diffusion-based SR methods on real-world LR inputs for $\times 4$ and $\times 8$ upscaling. The two LR images are taken from the RealSet65 dataset. For each image, a green region is selected for $\times 4$ upscaling and a blue region for $\times 8$ upscaling. FlowMapSR-2 consistently achieves the best balance between reconstruction faithfulness and photorealism.
  • Figure 5: Qualitative comparison of FlowMapSR with varying numbers of inference steps (NFE) for $\times 4$ (rows 1–2) and $\times 8$ (rows 3–4) upscaling. The LR inputs are drawn from the DIV2K-Val dataset. Increasing NFE improves perceptual quality, while fewer steps yield higher faithfulness to the reference, illustrating the trade-off quantified in \ref{['table:ablation_num_steps']}.
  • ...and 13 more figures

Theorems & Definitions (1)

  • proof