A short timeline
| Year | Model | Why it mattered |
|---|---|---|
| 2014 | SRCNN (Dong et al.) | First deep CNN for super-resolution |
| 2017 | SRGAN (Ledig et al.) | Adversarial and perceptual losses: sharp, realistic texture instead of blur |
| 2018 | ESRGAN (Wang et al.) | Improved SRGAN; became the community’s workhorse and model zoo |
| 2021 | Real-ESRGAN (Wang et al.) | Trained on realistic, messy degradations; works on real photos and JPEGs |
| 2021 | SwinIR (Liang et al.) | Swin Transformer for restoration; strong fidelity |
| 2021 | SR3 (Saharia et al.) | Showed diffusion models can do high-quality super-resolution |
| 2023 | StableSR (Wang et al.) | Re-used a pretrained Stable Diffusion model as a prior for upscaling |
| 2024 | SUPIR (Yu et al.) | SDXL-based, text-guided restoration at large scale |
| 2025 | SeedVR / SeedVR2 (ByteDance) | Diffusion-transformer restoration for video and images; SeedVR2 does it in one step |
Years refer to the first public paper or preprint.
GAN-trained CNNs (ESRGAN family)
A convolutional network upscales the image in one pass. During training a second network, the discriminator, tries to tell its outputs from real photos, which pushes it towards sharp, realistic texture. Strengths: fast, cheap, deterministic, runs locally, huge library of specialised models (anime, faces, textures, pixel art). Weaknesses: can look painterly or over-sharpened on very small inputs; limited “imagination” for large factors. Used in: Upscayl, chaiNNer setups, many web “standard” modes.
Transformers (SwinIR and successors)
Attention lets the model use context from further across the image, which helps with repeated structures such as windows, text and fabric. Strengths: high fidelity, good at structure. Weaknesses: heavier than CNNs, still essentially reconstructive. Often found behind “high fidelity” modes.
Diffusion upscalers
The model starts from noise and repeatedly denoises towards an image that is consistent with the low-resolution input, often steered by a text prompt and borrowing the “visual knowledge” of a large image generator. Strengths: rich, believable texture at large factors; prompts let you steer materials. Weaknesses: slower and costlier, non-deterministic, and the most likely to change faces, text and small objects. Used in: creative upscalers such as Magnific, Krea’s enhancer and Clarity, plus SUPIR and SeedVR-based tools.
So which should you use?
| If you need… | Choose |
|---|---|
| Speed, batches, privacy, predictable output | GAN / CNN (Real-ESRGAN family) |
| Maximum faithfulness on photos and text | Transformer or a “high fidelity” mode |
| Rich new texture, AI art, large factors | Diffusion |
| Both | Faithful base plus masked diffusion pass (how) |
Several of these models can be tried side by side in a browser. Upsampler’s upscaler model comparison lists the ones it runs.
FAQ
Is Real-ESRGAN still worth using in 2026?
Yes. It’s fast, free, runs on modest GPUs, and is predictable, which is what you want for photos, batches and anything that must stay faithful. Diffusion upscalers win on rich texture at a much higher computational cost.
Why are diffusion upscalers slower?
A GAN or transformer upscaler produces its output in one forward pass. Classic diffusion models refine the image over many denoising steps, often on overlapping tiles. One-step distilled models such as SeedVR2 narrow the gap.
Which family does Magnific use?
Magnific hasn’t published its architecture in detail. Its behaviour (prompts, creativity and resemblance controls, generated texture) is characteristic of diffusion-based upscaling.