Research map / Image generation / Super-resolution
APISR: Anime Production Inspired Real-World Anime Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes APISR, an image-based real-world anime super-resolution method trained on a curated I-frame image dataset (API) with a prediction-oriented compression degradation model, hand-drawn line enhancement for pseudo-GT, and a balanced twin perceptual loss combining anime (ResNet50) and photorealistic (VGG) features.
Problem
Real-world anime super-resolution methods rely on video networks/datasets and photorealistic-domain techniques, leading to faint/distorted hand-drawn lines and unwanted color artifacts in GAN-based outputs.
Approach
Exploit anime production traits: select least-compressed I-frames ranked by image complexity assessment to build an image dataset, simulate video compression on single frames via prediction-oriented compression, enhance hand-drawn lines in pseudo-GT, and use a balanced twin perceptual loss mixing an anime-trained ResNet50 with VGG features.
Key results
- On AVC-RealLQ at 4x, APISR (1.03M params) achieves NIQE 6.719 vs 8.109 for AnimeSR and MANIQA 0.514 vs 0.462, CLIPIQA 0.711 vs 0.567 for VQD-SR.
- Uses only 13.3% and 25% of the training sample complexity of AnimeSR and VQD-SR respectively.
- Ablation: ICA-based I-frame selection plus 720P rescale improves NIQE from 7.681 (AVC-Train) to 6.719.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use APISR instead of VQD-SR | Anime super-resolution · AVC-RealLQ | NIQE: 8.202 -> 6.719 | Helped |
| Use APISR instead of VQD-SR | Anime super-resolution · AVC-RealLQ | MANIQA: 0.464 -> 0.514 | Helped |
| Use APISR instead of VQD-SR | Anime super-resolution · AVC-RealLQ | CLIPIQA: 0.567 -> 0.711 | Helped |
| Use API instead of AVC-Train for training (ablation) | APISR · AVC-RealLQ | NIQE: 7.681 -> 6.719 | Helped |
| Use API instead of randomly selected training images (ablation) | APISR · AVC-RealLQ | NIQE: 8.006 -> 6.719 | Helped |
| Select I-Frames using ICA instead of IQA (ablation) | APISR training dataset · AVC-RealLQ | NIQE: 7.876 -> 6.912 | Helped |
| Add 720P rescaling to I-Frame + ICA selection (ablation) | APISR training dataset · AVC-RealLQ | NIQE: 6.912 -> 6.719 | Helped |
| Add prediction-oriented compression instead of High-Order degradation (ablation) | APISR degradation model · AVC-RealLQ | MANIQA: 0.483 -> 0.506 | Helped |
| Add prediction-oriented compression instead of Random Order degradation (ablation) | APISR degradation model · AVC-RealLQ | MANIQA: 0.491 -> 0.506 | Helped |
| Add shuffled resize to prediction-oriented compression (ablation) | APISR degradation model · AVC-RealLQ | NIQE: 7.133 -> 6.719 | Helped |
| Add hand-drawn line enhancement and twin perceptual loss (ablation) | APISR · AVC-RealLQ | CLIPIQA: 0.689 -> 0.711 | Helped |
| Add twin perceptual loss after hand-drawn line enhancement (ablation) | APISR · AVC-RealLQ | NIQE: 7.182 -> 6.835 | Helped |
Surprising findings
- Single-frame compression with a prediction-oriented codec can synthesize video-compression artifacts, so video networks and datasets are unnecessary for anime SR.
- IQA metrics favor simple monotonous scenes and penalize CGI-heavy frames, whereas image complexity assessment reverses this preference.
Limitations
- Anime-trained ResNet perceptual loss alone yields poor visuals due to Danbooru bias toward character faces and simple illustrations, so VGG is needed as an auxiliary.
- Evaluation relies only on no-reference metrics on a single benchmark (AVC-RealLQ).
Experimental setup
Datasets: API (Anime Production-oriented Image) dataset Danbooru AVC-Train
Benchmarks: AVC-RealLQ
Metrics: NIQE MANIQA CLIPIQA parameter count
Baselines: Real-ESRGAN BSRGAN RealBasicVSR AnimeSR VQD-SR
Builds on: Real-ESRGAN AnimeSR VQD-SR GRL BSRGAN
Closest papers by meaning
- SSIMBaD: Sigma Scaling with SSIM-Guided Balanced Diffusion for AnimeFace Colorization (NeurIPS 2025 · similarity 0.83)
- StdGEN: Semantic-Decomposed 3D Character Generation from Single Images (CVPR 2025 · similarity 0.82)
- Alchemist: Turning Public Text-to-Image Data into Generative Gold (NeurIPS 2025 · similarity 0.81)
- PBR-SR: Mesh PBR Texture Super Resolution from 2D Image Priors (NeurIPS 2025 · similarity 0.81)
- Goku: Flow Based Video Generative Foundation Models (CVPR 2025 · similarity 0.81)
- PhysAnimator: Physics-Guided Generative Cartoon Animation (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.