Przejdź do treści
Image Generator Parameters / AI Upscaling (Scale and Creativity)
Universal parameters

AI Upscaling (Scale and Creativity)

AI upscaling enlarges a finished image, controlled primarily by scale factor and creativity settings. A low creativity setting sharpens the original canvas while preserving every detail faithfully. Higher creativity grants the model freedom to invent brand-new micro-textures, which can elevate illustrations but may alter delicate features.

Also known as: upscale, upscale_factor, creativity, resemblance, Subtle, Creative

Side-by-side test results

AI Upscaling (Scale and Creativity) - x2, creativity 0.1
AI Upscaling (Scale and Creativity) - x2, creativity 0.6
x2, creativity 0.1x2, creativity 0.6

Drag the slider to compare. Left: x2, creativity 0.1. Right: x2, creativity 0.6.

AI Upscaling (Scale and Creativity) - original 512 x 512
original 512 x 512

Model: Clarity Upscaler (przez API fal.ai). Settings: input image 512 x 512 from resolution entry, upscale_factor 2, seed 1411. Date: 14.11.2026.
Prompt: domyślny prompt upscalera

Both test outputs measured 1024 px and looked notably sharper than the original file. At a creativity value of 0.6, the upscaler hallucinated crisper wood grain and surface textures, but it also subtly altered the pattern on the cup. Dropping creativity to 0.1 preserved the original cup design completely.

Syntax across generators

GeneratorParameterValuesDefault
Midjourney*Upscale (Subtle) / Upscale (Creative)Subtle: minimalne zmiany; Creative: dodaje detale, potrafi poprawić dlonie/miny2x-
Ideogramupscale endpoints (Topaz Bloom 2, Redefine, Standard V2, Text Refine, Wonder 3.5, Wonder 4, Nano Banana Pro, auto)Parameters 'resemblance'/'detail' (old upscale v1) are not present in API v2upscale_factor x2 (default)/4x/8x, result max. 8192 px per side; Wonder 4: 2x/4x, max. 100 MP; creativity 1-9 (def. 3) in Bloom 2/Redefine; strength 0.01-1 in Standard V2/Text Refine; Nano Banana Pro: resolution_tier 1K/2K/4K (def. 2K)x2
Imagen (retired)upscaleConfig.upscaleFactor (imagen-4.0-upscale-preview)Model in previewx2, x3, x4; result max. 17 MP-
Stability AI (SD 3.5)Upscale Fast / Conservative / Creative (creativity)Creative for heavily degraded images, not for photos >= 1 MPFast: 4x; Conservative: 20-40x up to 4 MP; Creative: 20-40x up to 4K, input 4096-1 048 576 px, creativity 0.1-0.5creativity 0.3
FLUXno separate upscaler in API (resolution up to 4k in FLUX 3)--
GPT Imagenone--

* Midjourney: docs.midjourney.com blocks automated scrapers, so values are sourced from search engine indexes. Please verify settings on the official Midjourney site before production use.

When to use it

Common mistake

A frequent pitfall is cranking up creativity on portraits or product mockups. The upscaler will gladly distort small typography, reshape facial features, or replace intricate packaging patterns with random decorative swirls. Another common mistake is chained upscaling, where running an image through multiple consecutive rounds stacks aggressive visual artifacts and unnatural contrast.

Ready-to-use example

Midjourney: click Upscale (Subtle) or Upscale (Creative)
Stability: /upscale/conservative or /upscale/creative with "creativity": 0.3
Imagen (Vertex): upscaleFactor x2 / x4
General definition (independent of generator settings): AI Upscaling (AI Image Enhancement) in the AI Glossary.

Frequently asked questions

What is the difference between Subtle and Creative upscaling in Midjourney?

Subtle upscales the image while keeping pixels and details as true to the original as possible. Creative re-interprets the canvas, introducing fresh micro-details and minor shifts in lighting or texture.

How does AI upscaling differ from traditional bicubic enlargement?

Traditional software stretches existing pixels and blurs the gaps, leading to soft edges. AI upscaling predicts and generates new realistic pixels based on training data, genuinely restoring perceived sharpness.

What does the resemblance slider do in AI upscalers?

Resemblance acts as the inverse of creativity: higher resemblance locks down the exact shapes and colors of the source image, while lower resemblance gives the model permission to improvise.

Related terms

🛒 Sklep Kinetyka CapCut Pro 460 zł/rok Lovable Pro 450 zł/rok Replit Core 119 zł/rok Zobacz wszystko →
× ‹ powiększenie ›