) and badly far above it, so the way to a large image is to generate at native size and enlarge afterwards. There are three ways to enlarge, and they are not interchangeable. Pixel

upscalers

are small networks (ESRGAN family, SwinIR, DAT) that sharpen while resizing; they are fast, faithful and add no new content. Hires-fix is a second diffusion pass at a larger size during generation that adds real detail at modest denoise. Tiled diffusion cuts a big image into overlapping tiles and runs

img2img

on each, which can invent plausible detail at 4K and beyond, guided by a Tile

ControlNet

so the tiles stay faithful.

Which one you need depends on the goal. For web and social use, a 2x pixel upscale of a clean render is enough and takes two seconds. For prints, products and wallpapers, a pixel upscale to 2x followed by a tiled diffusion pass at low denoise gives the detail people expect from "4K AI art". Hires-fix sits in between and is the default in

Forge

and A1111 for a reason: it fixes the mushy 1024 face without a second workflow.

The Upscalers family on this site lists the pixel upscalers with their intended use (photo, anime, faces, text) and the diffusion-based enhancers with their base family. The cloud partners run the same tools:

PirateDiffusion

has /upscale and /enhance commands,

BitVector

an upscale button next to each result.

Reference

Name	Type	What it is
4x-UltraSharp	pixel upscaler, ESRGAN	The default for photos and renders: sharp, slightly crunchy; downscale to 2x afterwards for a natural look.
Real-ESRGAN x4plus / x4plus-anime	pixel upscaler	Robust general purpose; the anime variant keeps lines clean on illustration.
4x-NMKD-Siax, 4x-Foolhardy-Remacri	pixel upscaler	Gentler alternatives to UltraSharp; Remacri for soft photos, Siax for detailed textures.
SwinIR / DAT / HAT	pixel upscaler, transformer	Higher fidelity, slower; best for photographs where ESRGAN invents texture.
Hires-fix	generation setting	Upscale by 1.5-2x with a pixel upscaler then denoise 0.3-0.5 at the new size. A1111/Forge checkbox; ComfyUI as a second KSampler.
Ultimate SD Upscale	tiled diffusion	Tiles of 1024 with 64 px overlap, denoise 0.2-0.35, Tile ControlNet on. Turns 2048 into 4096 with new detail.
SUPIR / Flux tiled / SeedVR	diffusion restorers	Heavy models that restore and enhance old or low-quality images; 16 GB+ and minutes per image.
Denoise	0.2-0.5	In any diffusion upscale, how much may change. 0.2 keeps everything; 0.5 redraws textures and may alter faces.

One-click upscale

Every image you make in BitVector Prism has an Upscale button that runs the right upscaler for the model; the result downloads at 2x or 4x. No model files to collect.

Try BitVector

Step by step

Generate at native size and fix composition and faces first (inpainting); upscaling amplifies flaws rather than fixing them.
Quick path: Upscale Image (using Model) with 4x-UltraSharp, then Image Scale to 2x of the original. Done in seconds; good for web.
Hires-fix path (Forge, A1111): enable Hires. fix, upscaler 4x-UltraSharp or Latent, upscale by 1.5-2, hires steps 15, denoise 0.4. In ComfyUI: Upscale Latent By 1.5 -> second KSampler at denoise 0.45.
Print path: pixel upscale to 2x, then Ultimate SD Upscale with the same
checkpoint
, tile 1024, padding 32, denoise 0.25, Tile ControlNet strength 0.6, seams fix "half tile". Expect several minutes.
Faces: after the upscale run a face detailer pass (ADetailer, FaceDetailer node) at denoise 0.35; large images make small faces again.
Sharpen lightly and export as PNG for prints, WebP or JPEG 92 for the web. Compare with the 1x render at 100 percent to catch invented artefacts.
On the cloud: reply to a PirateDiffusion result with /upscale (pixel) or /enhance (diffusion); in BitVector press Upscale on the image card.

Examples

ComfyUI quick 2x

VAE Decode -> Upscale Image (using Model: 4x-UltraSharp) -> Image Scale (0.5, lanczos) -> Save Image

Forge hires-fix settings

Hires. fix: upscaler 4x-UltraSharp | upscale by 1.5 | hires steps 15 | denoising strength 0.4 | same seed

Ultimate SD Upscale for a print

Tips

Upscale the final, not every draft; a 4x pass on a 1024 render is 16 million pixels and takes real time.
ESRGAN models are trained for 4x; running them at 2x means 4x then downscale, which is why the output looks oversharp when you skip the downscale.
Anime and flat illustration need the anime variants or the lines double; photos need SwinIR, DAT or Remacri or skin turns to plastic.
In tiled diffusion keep the prompt short and generic ("highly detailed, sharp focus, natural skin") so no tile invents a new subject.
Latent upscale in hires-fix needs denoise 0.5+ to be sharp; pixel upscalers in hires-fix work at 0.3-0.4 and keep the image closer to the original.
The Upscalers family page sorts by hot; the top three cover 90 percent of needs.

VAE Decode (Tiled) and smaller tiles.

/upscale and /enhance from the chat

Reply to any PirateDiffusion render with /upscale for a clean 4x or /enhance for a tiled diffusion pass with new detail. Server GPUs do the heavy lifting; unlimited passes on a fixed price.