Upscaling
last edited
Upscaling is the process of increasing image resolution after generation. Diffusion models have a native resolution they were trained at, and generating much larger makes compositions duplicate and smear. The standard workflow therefore generates at native size and upscales afterward.
Upscaling approaches
Two families of tools handle the task. Dedicated super-resolution networks, such as the ESRGAN lineage and its anime-specialized variants, sharpen and enlarge deterministically. Diffusion-based approaches re-generate the image at higher resolution and add detail that never existed in the source.
The classic hires fix does the latter in one pass. It upscales the image, then runs img2img at low denoise. The diffusion model invents plausible fine texture while preserving the composition.
Tiled upscaling and delivery
Tiled upscaling reaches poster-size outputs by processing overlapping patches. ControlNet tile guidance keeps each patch faithful to the source. This avoids the patchwork seams and hallucinated micro-objects of early attempts.
Print work and client delivery make this stage routine. Aggressive upscalers give images a crunchy, over-sharpened look. Diffusion detail passes can subtly redraw faces, and a fixed seed helps reproducibility. The choice of upscale model is as aesthetic a decision as the choice of checkpoint. See core techniques.
Tool landscape
SUPIR pairs a large vision model with diffusion priors for photorealistic restoration. SeedVR2 is a ByteDance video restoration model. Its repository lists it as an ICLR paper, alongside the earlier SeedVR.
Topaz Gigapixel is now on a 1.x version line. Release 1.3.4 shipped a detail-recovery model in August 2026. Release 1.5.0 arrived in November 2026. It adds a cloud-only model, with a local version announced as coming soon. Magnific offers slider-controlled creative upscaling.
Upscayl is a free desktop app built on Real-ESRGAN-architecture models. Its latest tagged release on GitHub is 2.15.
Sources
- Real-ESRGAN paper on practical restoration of real-world images
- SwinIR paper on transformer-based image restoration
- SUPIR on GitHub, the scaling-up restoration project
- SeedVR repository from ByteDance covering SeedVR2
- Topaz Gigapixel 1.5.0 release notes
- Upscayl releases on GitHub
- Magnific creative upscaling service