Remacri
- Skin, foliage, fabric and fine texture
- Portrait
UPSCAYL MODEL COMPARISON GUIDE
The best model depends on what is inside the image—not which preview looks sharpest.
A natural photograph, anime illustration, product image and old family photo contain different edges and textures. Use controlled testing to choose a believable result.
Model names, availability and licences can change. Verify the current release and model terms before production or commercial use.
SAME SOURCE, DIFFERENT PRIORITIES
Switch between representative image types. These examples explain what to inspect; they are not benchmark claims.

Inspect skin, hair, foliage, fabric and repeated background texture before approving the result.
VIDEO TUTORIAL
Use one source, identical scale and export settings, then compare matched crops and reject outputs with inaccurate detail.
MODEL SELECTION GUIDE
Use this table when you need a quick starting point.
These are starting points, not guaranteed winners. Test at least two models on important images.
If the result must be used commercially, verify the current license attached to the selected model. Do not assume that every model bundled with an open-source application has identical commercial-use terms.
View Side-by-Side Examples →MODEL SELECTION GUIDE
An Upscayl model is a trained image-processing system that estimates how a larger version of a low-resolution image could look.
Each model has learned different visual patterns. That affects how it interprets:
The model does not retrieve a hidden high-resolution original. It predicts plausible additional pixels from the information available in the source.
That distinction matters. A visually impressive result can contain details that were not present in the original scene.
Two models can produce files with identical pixel dimensions but visibly different results.
One may:
Another may:
Neither is automatically better. The correct choice depends on whether the result remains useful and believable at its final display size.
MODEL SELECTION GUIDE
The following guidance describes practical starting uses. Exact model availability and labels may change between Upscayl releases.
Do not publish these tendencies as benchmark results. Add real, controlled examples from the current application version before making performance claims.
MODEL SELECTION GUIDE
Upscayl Standard is a practical first test for mixed images when no specialized model is an obvious fit.
It can be useful for:
Use Standard as a baseline. Process the source with Standard first, then compare it with one specialized model.
Check whether:
Standard may be the best option when a specialized model produces a more dramatic but less believable result.
Upscayl Lite can be treated as a lighter starting option for preliminary tests or images that do not require aggressive reconstruction.
It may suit:
A model named “Lite” should not be assumed to be inferior. A lighter treatment can preserve an image more naturally when stronger reconstruction is unnecessary.
Compare Lite with Standard for:
Record actual processing time on your own hardware rather than publishing a universal speed claim.
High Fidelity is a strong starting candidate for natural photography and other realistic imagery. Upscayl identifies HFA2k as the model included under this name.
Consider it for:
The name does not mean that every generated detail is historically or photographically accurate. It remains an AI reconstruction.
Pay particular attention to:
If the output looks cleaner but subtly changes facial identity or material texture, compare it with Standard or Remacri.
Remacri is a useful comparison model for photographs, old images, and sources containing compression or mixed texture.
Consider testing it for:
Remacri may produce a balanced result where an edge-focused model looks too sharp.
Check:
For old photos, remember that upscaling is only one part of restoration. Scratches, stains, tears, fading, and missing regions may require separate manual work.
Ultramix Balanced is intended to balance detail and sharpness in natural imagery.
Potential uses include:
Its stronger definition may help an image feel clearer at a larger size. It can also make certain textures appear more processed.
Review:
The current Upscayl model list identifies Ultramix as noncommercial. Verify its current license before using it in paid work, client projects, advertising, ecommerce, or other commercial contexts.
Ultrasharp emphasizes edge definition and visible sharpness.
It may be useful for:
It is not automatically the best model for soft portraits, delicate film photographs, or images with heavy existing sharpening.
Look for:
Evaluate the image at its actual delivery size. An output that looks impressive at 200% zoom may look unnatural in a website layout or print.
The current model list identifies Ultrasharp as noncommercial. Confirm the applicable license before commercial use.
Digital Art is the logical starting point for images built from illustration lines, flat colors, and synthetic shapes.
Use it for:
It may preserve graphic boundaries more effectively than a photography-oriented model.
Check:
Do not use Digital Art automatically for every AI-generated image. Photorealistic AI artwork may respond better to High Fidelity, Standard, or Remacri.
MODEL SELECTION GUIDE
Start with High Fidelity and compare it with Standard or Remacri.
Natural photography contains irregular, subtle texture. The goal is usually to improve usable resolution without turning skin, foliage, stone, or clouds into synthetic patterns.
Begin with High Fidelity, then compare Standard or Remacri.
Choose the result that preserves identity—not the result with the strongest skin texture.
Inspect both eyes independently. AI systems can create small differences between them, especially when the original face occupies few pixels.
If face enhancement is available, compare:
Do not assume the enhanced face is more accurate.
Start with Remacri or High Fidelity.
Old photographs may contain grain, fading, scratches, dust, blur, or scanning artifacts. A model can interpret those defects as texture.
Use a conservative sequence:
Scan the original at a suitable resolution.
Preserve an untouched master.
Correct basic exposure and color carefully.
Test two upscaling models.
Inspect faces and clothing.
Complete manual restoration separately.
Start with Standard or High Fidelity. Compare Remacri when the source has compression artifacts.
Product accuracy is more important than dramatic detail.
Inspect:
Do not publish an output that invents texture or changes a material’s appearance.
Start with Digital Art.
Compare another model only when the artwork contains painted texture, realistic lighting, or photographic elements.
Inspect:
Choose the starting model based on the visual style:
Digital Art for graphic or illustrated images.
High Fidelity for realistic imagery.
Remacri for mixed or compressed outputs.
Standard when the image combines several styles.
Upscaling can strengthen existing AI defects. Inspect hands, eyes, jewelry, typography, and repeated decoration before approving the result.
Begin with High Fidelity or Standard. Test Ultrasharp only when stronger edge definition is useful and the intended use complies with its license.
Inspect:
Architecture makes small geometric distortions easier to notice than organic photography.
MODEL SELECTION GUIDE
A useful Upscayl model comparison changes only one variable: the model.
Use an image containing the details that matter for your intended workflow.
A photography test might include:
An illustration test might include:
Document:
Use the same:
Change only the model.
Do not process every available model when the image type makes some choices irrelevant.
For example:
Compare High Fidelity, Standard, and Remacri for a portrait.
Compare Digital Art and one realistic model for an illustration.
Compare Standard, High Fidelity, and an edge-focused model for architecture.
Create identical crops from every output.
Useful crops include:
Show the original beside every crop.
Inspect at:
Pixel-level sharpness is only one part of quality. The image must remain believable at its final size.
Use a test table like this:
Suitable for final use?
Do not fill this table with estimated data. Publish results only after testing the named version and hardware.
View Model Examples →MODEL SELECTION GUIDE
Look for changed identity, uneven eyes, unnatural teeth, plastic skin, altered glasses, or duplicated facial details.
Review hair, fur, fabric, grass, stone, and wood. Watch for repeated patterns and texture that appears painted onto the surface.
Inspect high-contrast boundaries for halos, jagged contours, doubled edges, and artificial outlines.
AI upscaling is not a dependable method for reconstructing unreadable text. Verify every letter, number, product label, and logo manually.
Check skies, shadows, illustrations, and studio backgrounds for banding or abrupt color changes.
A model may magnify compression, noise, blur, dust, and previous sharpening. Compare every suspicious area with the source.
Check product colors, skin tones, neutral grays, and saturated areas. A sharper result is not acceptable if it changes important colors.
MODEL SELECTION GUIDE
The model and enlargement scale solve different parts of the task.
The model affects how new pixels are estimated.
The scale determines the output dimensions.
The export settings affect compression and file size.
The source limits the reliable detail available.
A 4× model does not mean every image should be delivered at four times its original width and height.
Start with the smallest output that meets the final requirement.
Use a moderate enlargement for:
A larger enlargement can help with:
Running a second enlargement creates a much larger file and another reconstruction pass. It may strengthen both useful detail and unwanted artifacts.
Compare:
Choose the simplest workflow that produces a usable result.
MODEL SELECTION GUIDE
Upscayl supports additional compatible models, but custom models should be treated as an advanced option.
Consider one when:
Built-in models consistently fail on a recurring image type.
You have a specialized anime, texture, restoration, or photography workflow.
The model format is compatible with Upscayl’s backend.
Its license permits your intended use.
You can test the model on representative files.
Before installing a custom model, verify:
Do not download model files from an unknown source simply because an example looks impressive.
A custom model can produce stronger results on one curated example and worse results on everyday images.
Test it using:
MODEL SELECTION GUIDE
Upscayl Desktop and Upscayl Cloud are separate workflows.
Cloud currently advertises additional models and higher-scale options, but model availability can change. Check the live interface before naming a Cloud model in permanent page copy.
Compare Desktop and Cloud →MODEL SELECTION GUIDE
Strong sharpness can hide halos, false texture, and changed facial features.
Choose the most useful and believable result.
If the model, scale, compression, and sharpening all change, you cannot tell which setting caused the difference.
Change one variable at a time.
Every model must be evaluated using the same area at the same display size.
Include difficult images with faces, foliage, text, compression, and fine lines.
Separate photographs, illustrations, screenshots, and products before batch processing.
Open-source software can bundle or support models with different licenses. Verify each model’s terms before commercial use.
AI-generated texture can look convincing without being accurate.
This is especially important for:
A useful comparison should include at least one weak output and explain why it was rejected.
MODEL SELECTION GUIDE
Upscayl makes it possible to process the same source with several model types using a consistent desktop workflow.
Its main advantages include:
The model selector is valuable because there is no universal model for every image. The best result comes from controlled testing and careful review.
MODEL SELECTION GUIDE
There is no universal best Upscayl model. High Fidelity is a sensible starting point for realistic images, Digital Art for illustrations, and Standard or Remacri for general and mixed photographic sources. Test at least two relevant models.
Start with High Fidelity. Compare it with Standard or Remacri, particularly when the source contains compression, older photographic texture, or mixed detail.
High Fidelity is a reasonable first test, but the winning output should be the one that best preserves identity. Inspect eyes, teeth, hair, skin, glasses, and facial contours.
Start with Digital Art. It is designed for illustration-like content and clean graphic boundaries. Compare it with another compatible art model if the image contains painted texture.
Test Remacri and High Fidelity. Old photographs vary greatly, so review facial identity, film grain, clothing, and damaged areas closely.
It may produce stronger edge definition, but stronger sharpness can introduce halos and artificial texture. It is not automatically the best choice for portraits or soft photography.
The current Upscayl model list presents Ultrasharp as noncommercial. Verify the current license and model information before using it for client, ecommerce, advertising, or other commercial work.
The current model list also presents Ultramix as noncommercial. Confirm the terms applying to the exact model version before commercial use.
No model can guarantee recovery of detail that was absent from the source. Upscayl’s documentation explains that its models estimate what additional details could look like.
Upscayl states that it is not intended to perform focus adjustment or fully deblur severely blurred images. A model may improve the appearance of some edges without correcting the underlying focus problem.
Models are trained on different image characteristics and may reconstruct surfaces differently. Export settings and source color information can also affect the result.
Use the same output format and compression setting for every comparison. A high-quality or lossless format can make model differences easier to evaluate.
Use the same scale for every model. Start with the smallest enlargement that meets your goal. Excessive scaling can make model artifacts more visible.
Upscayl supports compatible custom models. Confirm the required architecture, files, scale, application compatibility, and license before installation.
Not necessarily. Desktop models are included with the installed application release, while Cloud maintains its own model catalog and features.
Not automatically. Speed and visual suitability are separate questions. Measure processing time and inspect output quality using the same source and hardware.
MODEL SELECTION GUIDE
Start with one representative image and two models suited to its content.
Keep the scale, output format, hardware, and source file unchanged. Compare the same face, texture, edge, and text crop before choosing the final output.
The right model is not the one with the most dramatic preview. It is the one that creates the most useful result without introducing distracting or inaccurate detail.
View Model Examples →Download Upscayl Desktop →Learn How to Upscale an Image →TEST TWO RELEVANT MODELS
Keep the source, scale, format and hardware unchanged. Select the model that produces the most useful and believable output—not the most dramatic preview.