In one paragraph
Faithful upscalers (ESRGAN, SwinIR, “standard” and “high fidelity” modes) reverse blur and compression and stay close to the input. Creative upscalers (diffusion-based, as popularised by Magnific) re-generate the image at high resolution, guided by the input. Use faithful when accuracy matters, creative when appearance matters, and a blend of the two when you need both.
Faithful (reconstructive) upscalers
These models are trained on pairs of images: a high-resolution original and a version that has been degraded on purpose (downscaled, blurred, noised, JPEG-compressed). They learn to undo that damage. Real-ESRGAN, for instance, trains on synthetic degradations designed to look like the messy real-world images people actually feed in. The result is sharper edges, cleaner gradients and fewer blocks, with no new objects. Faithful models are predictable: the same input gives the same output.
Use them for: real people, products and packaging, screenshots and UI, documents, architecture, and anything where a viewer would feel misled by invented detail.
Creative (generative) upscalers
Diffusion upscalers start from the low-resolution image, add noise, and let an image-generation model “denoise” it back at a higher resolution, guided by the input and often by a text prompt. Because the model is generating, it can add pores, fabric weave, leaf veins and brick texture that were never in the source. The common controls are:
| Slider (names vary) | What it changes | Start at |
|---|---|---|
| Creativity / denoise strength | How much the model may re-imagine | Low to medium for photos; higher for AI art |
| Resemblance / fidelity / structure | How tightly the output follows the input’s shapes | High for faces and products |
| HDR / detail | Local contrast and micro-detail | Moderate. High values look over-processed |
| Prompt | What the model thinks it’s looking at | Describe materials (“weathered oak, linen”) rather than moods |
Use them for: AI renders, concept art, game assets, landscapes, and stylised work where richer texture is welcome.
Which to use for which image
| Image | Use | Why |
|---|---|---|
| AI art, concept art | Creative | Invented detail is a feature |
| Portraits of real people | Faithful, or creative at low settings | Avoids changing who they are |
| Products, packaging, screenshots | Faithful | Creative models garble text and logos |
| Landscapes | Either | Generated texture usually looks natural |
| Old family photos | Restoration, then faithful | Keeps the record accurate |
| Anime, line art | Anime-tuned faithful | Keeps flat colour flat |
The blend workflow professionals use
- Faithful pass to the final size. This is your safe base.
- Creative pass at low to medium strength from the same source, at the same size.
- Stack both in an editor, creative on top, and mask it out of faces, text, logos and hands.
- Check at 100%, then export once, as PNG or high-quality JPEG.
To decide objectively which pass you prefer, compare them blind with the A/B tool. Upsampler has published a longer creative vs precise comparison with example images.
FAQ
Is Magnific a creative or a faithful upscaler?
Magnific is best known for creative upscaling, with sliders such as Creativity, HDR and Resemblance that control how far the result may drift from the input. Its pricing page also lists a “Magnific Precision” model.
What does the “creativity” slider actually do?
In diffusion upscalers it roughly sets how much noise is added before the model re-generates the image, or how weakly the output is tied to the input. Low values behave like a strong sharpener. High values repaint the image.
Can I combine both?
Yes, and it’s often the best workflow. Make a faithful upscale to the target size, run a light creative pass, then blend the two with a layer mask so faces and text come from the faithful version.