Efficient INT8 Single-Image Super-Resolution via Deployment-Aware Quantization and Teacher-Guided Training

Conference: CVPR 2026
arXiv: 2604.20291
Code: None
Area: Image Restoration / Super-Resolution / Model Quantization
Keywords: INT8 Quantization, Single-Image Super-Resolution, Structural Re-parameterization, Knowledge Distillation, Edge Deployment

TL;DR

For \(\times 3\) single-image super-resolution (SISR) on mobile NPUs, this paper utilizes a deployment-oriented pipeline featuring a "LR-space MobileOne re-parameterized backbone + three-stage teacher-guided training + fusion-before-QAT," achieving INT8 29.79 dB / 0.8634 SSIM with 82K parameters and a final score of 1.8 in the MAI 2026 Quantized SR Challenge.

Background & Motivation

Background: SISR has recently improved accuracy by increasing capacity—EDSR deepens residual backbones, RCAN use residuals-in-residuals + channel attention, and SwinIR/HAT introduce window attention for long-range dependencies. These models manifest strong fidelity but are increasingly large and difficult to compress.

Limitations of Prior Work: Deploying SR models on mobile NPUs for INT8 execution faces three major obstacles. First, quantization sensitivity: SR quality is measured at the pixel level; activation ranges, rounding errors, and training-deployment inconsistencies directly manifest as visible blurring and artifacts, making it much harder to quantize than high-level vision tasks. Second, insufficient capacity: compact models inherently lack the ability to restore complex textures and long-range structures. Third, structural mismatch: re-parameterized backbones are multi-branch during training and fused into single-branch for inference; performing Quantization-Aware Training (QAT) directly on the multi-branch structure leads to unpredictable accumulation of quantization errors and accuracy collapse after branch fusion.

Key Challenge: There is a trade-off between reconstruction fidelity, model compactness, and low-bit robustness. Most works optimize only one aspect, treating quantization as a post-processing "conversion" rather than a primary design goal.

Goal: To develop a compact \(\times 3\) SR pipeline capable of real-world INT8 execution on mobile NPUs, ensuring strict alignment between the training optimization graph and the deployment integer graph.

Key Insight: Rather than inventing new SR operators, "architectural design + supervisory signals + deployment consistency" are treated as a joint optimization problem. Computation is kept in low-resolution (LR) space to save power, a Mamba teacher compensates for capacity shortfalls, and QAT is applied directly to the fused deployment graph to eliminate mismatches.

Core Idea: Utilize a "LR-space re-parameterized backbone + teacher-guided multi-stage fidelity training + deploy-before-QAT" triplet to enable an 82K parameter model to approach FP32 quality under INT8.

Method

Overall Architecture

The input is a low-resolution RGB image \(x\in\mathbb{R}^{H\times W\times 3}\), and the output is a \(\times 3\) magnified \(\hat{y}\in\mathbb{R}^{3H\times 3W\times 3}\). The method follows two paths: the Student Network performs compact inference (extract–refine–upsample) in LR space, while the Three-stage Training Pipeline progressively aligns floating-point fidelity with INT8 deployment constraints. The student uses a stem to project LR images into feature space, processes them through 8 MobileOne-style blocks in the LR domain, employs global skip connections for structural preservation, and finally reconstructs via a PixelShuffle head. Training proceeds from Stage 1 (L1 initialization) to Stage 2 (Charbonnier + DCT + Teacher Distillation) and finally Stage 3 (QAT on the fused deployment graph), eventually exporting as TFLite INT8.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["LR Input x"] --> B["LR-space Re-parameterized Backbone<br/>stem + 8×MobileOne blocks<br/>+ Global Skip + PixelShuffle"]
    B --> C["Stage 1: L1 Foundation<br/>Learn Stable Spatial Mapping"]
    C --> D["Stage 2: Teacher-Guided Fidelity<br/>Charbonnier+DCT+Confidence Distillation"]
    D -->|BN Recalibration + Branch Fusion| E["Stage 3: Deploy-before-QAT<br/>Three-Stage QAT Curriculum"]
    E --> F["Export TFLite<br/>INT8 NHWC Deployment Graph"]

Key Designs

1. LR-space Re-parameterized Backbone: Multi-branch for training, single-branch for inference

The challenge is for compact SR to save power while maintaining expressivity. This paper restricts most computation to LR space—the stem uses a \(3\times 3\) convolution to project input to \(C=32\) channels, and \(N=8\) MobileOne blocks refine features. Upscaling happens at the end via PixelShuffle; thus, the computational load scales with \(H\times W\) rather than \(3H\times 3W\), making it mobile-friendly. Training uses multi-branch structures (\(4 \times 3\times 3\) convs, \(1\times 1\) conv, identity) with BN and ReLU:

\[B(f)=\sigma\Big(\sum_{i=1}^{5}\mathrm{BN}_i(\mathrm{Conv}^{(i)}_{3\times 3}(f))+\mathrm{BN}_{1\times 1}(\mathrm{Conv}_{1\times 1}(f))+\mathrm{BN}_{\mathrm{id}}(f)\Big)\]

After training, branches are folded using the BN folding formula \(\widetilde{W}=\frac{\gamma}{\sqrt{\sigma^2+\epsilon}}W,\ \widetilde{b}=\beta+\frac{\gamma}{\sqrt{\sigma^2+\epsilon}}(b-\mu)\) into a single \(3\times 3\) convolution. A global skip connection \(f=f_N+f_0\) is added before the PixelShuffle head. Ablations show MobileOne blocks are more robust to quantization than RepConv or RepDW.

2. Teacher-Guided Three-stage Training: Compensating for texture shortfalls via Mamba Teacher

Compact models struggle with fine textures and long-range structures; hence, a pre-trained MambaIRv2Light \(\times 3\) teacher is used for output-level distillation. Training proceeds progressively: Stage 1 uses L1 loss; Stage 2/3 switch to Charbonnier loss \(\mathcal{L}_{\mathrm{char}}=\frac{1}{N}\sum_i\sqrt{(\hat{y}^i_{01}-y^i_{01})^2+\epsilon^2}\) (\(\epsilon=10^{-3}\)) and add DCT frequency supervision \(\mathcal{L}_{\mathrm{DCT}}=\|D(\hat{y}_{01})-D(y_{01})\|_1\). Distillation uses confidence weighting: pixel-wise weights \(w(p)=\mathrm{clip}(\exp(-\gamma e(p)),w_{\min},w_{\max})\) are calculated based on teacher error \(e(p)=\frac{1}{3}\sum_c|t_c(p)-y_{01,c}(p)|\). The distillation loss is \(\mathcal{L}_{\mathrm{KD}}=\frac{1}{N}\sum_p w(p)|\hat{y}_{01}(p)-t(p)|\). Stage 3 with the teacher improves INT8 from 29.91 dB to 30.00 dB.

3. Deploy-before-QAT: Ensuring alignment between fake-quant and real INT8 graphs

This is the core solution for re-parameterization/quantization mismatch. QAT directly on multi-branch graphs cause unpredictable errors after fusion. This paper reverses the process: the network is collapsed into deployment form before QAT initialization. After forward-only BN recalibration (64 batches), branches are fused into single \(3\times 3\) convs. QAT operators are then inserted into this fused graph using PyTorch FX graph-mode. QAT follows a three-stage curriculum: Epoch 0–30 activates observers to calibrate scale/zero-point; Epoch 30–90 freezes grids and fine-tunes weights via STE; Epoch 90–150 freezes fake-quant nodes for final convergence. Weights are clipped \(W\leftarrow\mathrm{clip}(W,W_{\min},W_{\max})\) to suppress outliers.

Key Experimental Results

Main Results

Dataset: DIV2K (800 train / 100 val / 100 test), task \(\times 3\) SR, RGB PSNR/SSIM evaluation.

MAI 2026 Challenge Leaderboard (Quantized 4K SR, Selected):

Team	FP32 PSNR/SSIM	INT8 PSNR/SSIM	NPU Latency(ms)	Final Score
AntYSP	30.04 / 0.8757	29.96 / 0.8729	4.33	21.8
AntSR	30.08 / 0.8764	29.98 / 0.8731	4.52	21.5
z6	N.A.	29.93 / 0.8699	5.49	16.5
IN2GM	29.88 / 0.8708	29.85 / 0.8701	46.5	1.7
AIO_MAI (Ours)	29.98 / 0.8730	29.79 / 0.8634	41.1	1.8

Comparison with baselines (Ours uses fixed-shape deployable INT8 TFLite, only 82K parameters):

Method	Type	PSNR (dB)	SSIM
Bicubic	FP32	28.26	0.828
FSRCNN	FP32	29.45	0.838
ABPN	INT8	30.15	0.852
Ours (Stage 2)	FP32	30.28	0.863
Ours (Deploy)	INT8	30.13	0.858

Ablation Study

Backbone Comparison (FP32 to Dynamic INT8 TFLite drop):

Block	FP32 PSNR/SSIM	INT8 PSNR/SSIM	ΔPSNR↓	ΔSSIM↓
RepConv	30.0897 / 0.855	29.8492 / 0.847	0.2405	0.008
RepDW	29.3583 / 0.838	28.7031 / 0.814	0.6552	0.024
MobileOne	30.1350 / 0.859	30.0003 / 0.856	0.1347	0.003

Stage 3 Teacher Supervision Ablation:

Config	Teacher	Precision	PSNR (dB)	SSIM
Stage 3 (Direct QAT)	✗	INT8	29.9114	0.853
Stage 3 (Ours Full)	✓	INT8	30.0003	0.856

Key Findings

• MobileOne is the optimal anti-quantization solution: It achieves the highest INT8 PSNR (30.0003 dB) and the smallest FP32→INT8 drop (0.1347 dB). RepDW suffers significant drops (0.6552 dB), indicating that aggressive factorization is detrimental to low-bit SR.
• Teacher supervision remains effective during final quantization: Adding the teacher during Stage 3 pulls the INT8 result above 30 dB, proving that strong supervision bridges the optimization gap.
• Quantization must be a primary design goal: Even minor drops during export can disqualify a model; "deploy-before-QAT" ensures near-lossless export.

Highlights & Insights

• "Deploy-before-QAT" directly addresses the fundamental mismatch: This work resolves the issue where "training graph \(\neq\) deployment graph" causes accumulated quantization errors. This train-deploy consistency is applicable to any re-parameterized low-bit scenario.
• Practical Confidence-Weighted Distillation: By weighting pixels according to teacher error, the student avoids learning blurred artifacts.
• DCT Supervision compensates for CNN high-frequency weaknesses: Explicitly constraining DCT coefficients targets SR-critical details more effectively than L1 loss alone.
• 82K Parameters matching ABPN: Approaching larger model quality within minimal parameter budgets validates that "careful coordination > scaling up."

Limitations & Future Work

• Absolute accuracy trails top-tier teams: The Final Score of 1.8 is significantly lower than the leader (21.8), primarily due to NPU latency (41.1ms vs 4.33ms).
• Validated on a single target platform: Evaluation was limited to the MAI 2026 NPU; generalization to MediaTek or Apple Neural Engine is unverified.
• Minor gains from teacher distillation: The 0.09 dB improvement may not justify the training cost of a Mamba teacher in non-challenge scenarios.
• Systematic integration over operator novelty: Innovation lies in the pipeline rather than new SR primitives.

Related Work & Insights

• vs QuantSR: While QuantSR focuses on quantization-aware calibration and clipping, this work emphasizes "deploy-before-QAT" consistency.
• vs MobileOne / RepVGG: Uses re-parameterization but specifically fills the gap where these methods clash with INT8 quantization.
• vs DVMSR: Both use Mamba teachers, but this work pushes distillation into real INT8 edge deployment with confidence weighting.

Rating

• Novelty: ⭐⭐⭐☆☆ No new operators; novelty lies in consistency insights + distillation combinations.
• Experimental Thoroughness: ⭐⭐⭐⭐☆ Includes challenge results, baselines, and dual ablations, though platform-limited.
• Writing Quality: ⭐⭐⭐⭐☆ Clear logic, well-documented formulas, and flow.
• Value: ⭐⭐⭐⭐☆ Practical deployment-aware ideas for INT8 SR; highly transferable for re-parameterized quantization.

Related Papers

• [CVPR 2026] IAFMNet: Information-Aware Feature Modulation for Efficient Super-Resolution
• [ICCV 2025] Outlier-Aware Post-Training Quantization for Image Super-Resolution
• [CVPR 2026] LightRR: A Lightweight Network for Single Image Reflection Removal
• [CVPR 2026] ExpoCM: Exposure-Aware One-Step Generative Single-Image HDR Reconstruction
• [CVPR 2026] PS-SR: Pseudo-Single-Step Video Super-Resolution via Speculative Diffusion