Upscale Models

One of the most popular 4x upscalers for photorealistic images. Produces crisp, sharp results with excellent detail preservation across all styles.

Official UltraSharp upload on Hugging Face by Kim2091. Available in both .safetensors and .pth formats for maximum compatibility.

The most downloaded upscaler on Civitai with 22K+ overwhelmingly positive reviews. A reliable, versatile 4x upscaler for all image types.

A highly popular general-purpose 4x upscaler by NMKD. Works best with realistic and photographic images with 1.2K+ overwhelmingly positive reviews.

Balanced 4x upscaler trained for 200K steps by NMKD. Delivers clean results across realistic and anime styles in SafeTensor format.

General-purpose upscaler by Kim2091 that balances sharpness, detail, and smoothness. Available in three variants: Smooth, Balanced, and Restore.

The best overall upscaler for most users — 9.2/10 quality with excellent real-world image handling. Trained on synthetic data for practical restoration.

The 2x variant of RealESRGAN for moderate upscaling. Ideal when 4x is too aggressive or you need faster processing with high-quality results.

Anime-optimized RealESRGAN with only 6 RRDB blocks — smaller (17 MB) and faster while delivering excellent anime upscaling with clean line preservation.

Ultra-compact 2.4 MB model for anime video upscaling. Improved background restoration, faithful colors, fewer artifacts, and better naturalness.

Dedicated anime upscaler by Kim2091 that produces crisp, detailed results optimized for linework and cel-shaded illustration styles.

Transformer-based upscaler using Shifted Window attention. Achieves state-of-the-art 9.7/10 quality for image SR, denoising, and JPEG artifact reduction.

Next-gen SwinV2-based image restoration model. Handles classical SR, compressed image SR, and real-world SR with improved training stability.

Blind super-resolution model trained on complex degradation combinations — blur, downsampling, noise, and JPEG compression. Excels on real-world photos.

Photo upscaler based on the DAT (Dual Attention Transformer) architecture. Handles JPEG compression, blur, and resize with 110K iterations of training.

Hybrid Attention Transformer upscaler combining channel and self-attention with overlapping cross-attention for superior photo restoration at 157 MB.

Kim2091's best model ever — based on DAT2 architecture. Handles realistic images, anime, cartoons, and artwork with a lighter RealPLKSR variant available.

The classic 2x Enhanced Super-Resolution GAN. A reliable choice for moderate upscaling when 4x is overkill — fast and efficient.

ESRGAN-based photo upscaler trained on the Nomos8k dataset. Specializes in JPEG compression removal and blur correction for real-world photographs.

Maximum 8x upscaling variant of RealESRGAN for extreme resolution enhancement. Useful for very low-resolution source images or print preparation.

ByteDance one-step video restoration model with 3B parameters. Upscales 540p to 4K with temporal consistency via diffusion and 3D U-Net architecture.

Full-size 7B parameter SeedVR2 for maximum quality video upscaling. Includes a "sharp" variant for enhanced detail — the state-of-the-art in video SR.

Real-time diffusion-based streaming video SR at ~17 FPS on A100. Three-stage distillation with locality-constrained sparse attention for 12x speedup.

ComfyUI-optimized FlashVSR with tiny, tiny-long, and full modes. No custom kernel compilation needed — uses Sparse_Sage attention for easy setup.

Semantic-aware diffusion upscaler that intelligently reconstructs missing details. The Quality variant prioritizes visual enhancement — outperforms Topaz AI.

The Fidelity variant of SUPIR that prioritizes accuracy to original content. Best for restoring degraded, compressed, or old photos while preserving details.

Diffusion + GAN hybrid that reduces output variability via non-uniform timestep learning. Maintains content consistency across upscaled images and video.

Stable Diffusion-based super-resolution with time-aware encoder and controllable feature wrapping. Trained on DF2K + OST for high-quality upscaling.

Two-stage blind restoration framework — removes degradations first, then regenerates details via Stable Diffusion with LAControlNet guidance.

GigaGAN-based 4x super-resolution by fal.ai. Image-conditioned upscaling with overlapped tiling to eliminate seam artifacts — 2.47 GB SafeTensor model.

CVPR 2024 anime-specific super-resolution. Analyzes anime production workflows to handle distorted hand-drawn lines and color artifacts with twin perceptual loss.

Ultra-lightweight 1.7 MB upscaler with only 792K parameters. Uses Omni Self-Attention to model pixel interactions across spatial and channel dimensions.

CVPR 2024 efficient SR with dense-residual connections that prevent spatial info loss. Outperforms HAT with fewer FLOPs — 14M params for the base model.