Ranking Image Upscalers
Image upscaling is simple to describe and hard to evaluate. Given a low-resolution image, an upscaler must produce a larger version that looks more detailed without changing what was actually there. A result can look sharp at first glance while quietly inventing letters, faces, textures, or architectural details the source never had.
That makes isolated demos a poor basis for comparison. Models need to be tested on the same source images, at the same target resolution, and inspected at the same scale. The Crafiq Image Upscale Arena does exactly that: it shows two anonymous results side by side and asks which one is better.
Early results put ByteDance's SeedVR2, Clarity AI's Crystal Upscaler, and Pruna AI's P-Image-Upscale at the front of the pack — though the evaluation is still young, and the ranking will shift as the sample grows.
How the arena works
Every test begins with the same 1024x1024 pixel source image. Each model upscales it to 4096x4096 pixels: four times the width and height, and sixteen times the total number of pixels. Keeping the task fixed makes the results directly comparable.
In each round, the arena shows two outputs for the same case with the model names hidden. A before-and-after slider lets you compare each 4K result directly with the original. Zoom and pan are synchronized across both sides, making it possible to inspect the same region in each image. This is important because the most revealing differences are often small: the edge of a label, the shape of an eye, a distant vehicle or a few pixels of text.
Once a preference has been submitted, the model names are revealed and the result contributes to the global ranking. Keeping the identities hidden until then helps reduce the influence of brand familiarity, release recency and expectations about a particular model family.
The test set covers different types of image, including dense handwritten text, product packaging, people, an archival photograph, a technical document, a software interface, and more. Each category exposes different weaknesses. A model that performs well on a clean portrait may be much less reliable when it encounters small text, repeated structures or objects represented by only a handful of pixels.
What the comparisons reveal
The crops below show the original image alongside results from ESRGAN, Google Upscaler, SeedVR2 and Crystal Upscaler. For a more detailed and interactive comparison of all the models and cases, see the Image Upscale Compare page.
The supermarket shelf concentrates several difficult elements into one small crop: a curved bottle, a logo, hazard symbols and a price label. ESRGAN strengthens edges and noise but recovers little readable structure. The newer models produce cleaner boundaries and more coherent-looking lettering. In this example, SeedVR2 and Crystal Upscaler are the strongest results. Both preserve the overall bottle and label structure well, although neither can know the missing characters with certainty. Plausible-looking text should therefore not be mistaken for verified text.
The archival street photograph presents a different challenge. The source contains only a blurred suggestion of the subject's face, yet Google Upscaler reconstructs a notably clear and coherent portrait. Among the models shown, it is the clear best result for this case. SeedVR2 and Crystal Upscaler remain softer and preserve more of the source photograph's grain and ambiguity, but recover less useful facial structure.
Even a convincing reconstruction still needs to be interpreted carefully. Much of the facial detail was not explicitly available in the low-resolution source and has therefore been inferred by the model. For a visual restoration that may be desirable; for historical or forensic use, it is important not to treat generated detail as evidence and clearly label any output as a reconstruction to avoid misinterpretation.
The city scene tests tiny objects and repeated geometry. The bus, cars, lane markings, railings and pedestrians occupy very few pixels in the source. Some models create an impression of detail by increasing local contrast, while others reconstruct shapes more aggressively and risk changing their form. Here, SeedVR2 and Crystal Upscaler again appear roughly tied for the best result. Both define the vehicles street layout and crosswalk more clearly than the other models.
These examples also show why there is no single universal definition of a good upscale. Product photography may benefit most from clean edges and attractive texture, while invented content is a serious drawback in documents, interfaces and archival material. No model wins every type of image: Google Upscaler is the clear standout in old grayscale photographs, while SeedVR2 and Crystal Upscaler are more consistent across the other cases.
The three images above are only a small sample. In the Image Upscale Arena, the same models can appear across other cases alongside many additional upscalers. On the Image Upscale Compare page, you can select the models and examples yourself and examine their outputs in more detail.
Where the ranking stands
The frontier graph below plots arena score against release date. It shows the broad progression from ESRGAN and GFPGAN through Google Upscaler to SeedVR2, while also making clear that a newer release is not automatically better. Models below the frontier may still be practical choices when speed, cost, licensing or a particular visual character matters more for a given workflow.
At this early stage, SeedVR2 leads, followed by Crystal Upscaler and P-Image-Upscale. The three leading models represent different practical trade-offs:
- SeedVR2 is an open-weights model from ByteDance, released in June 2025. In the current benchmark configuration it costs about $0.016 per image and takes 18.52 seconds.
- Crystal Upscaler is from Clarity AI and was released in August 2025. Its current measured time is 16.91 seconds, with a cost of about $0.20 per image.
- P-Image-Upscale is from Pruna AI and was released in April 2026. It is the fastest of the current top three at 9.65 seconds, with a cost of about $0.02 per image.
These cost and runtime figures provide useful benchmark context, but they can vary with the provider, hardware and pricing at the time of measurement. They should not be read as quality indicators either: the arena score reflects visual preference, while cost, speed and release date describe the practical trade-offs surrounding that result.
The full leaderboard covers a much wider range of models. GFPGAN, for example, is faster and less expensive than the leading group, while Clarity Pro Upscaler takes substantially longer and costs more. Those differences can be decisive in a high-volume workflow, but they do not replace close inspection of the actual output.
Because the arena is new, the score gaps and confidence ranges remain provisional. The current leaders are a snapshot rather than a permanent conclusion. As the arena collects broader coverage across cases and model pairings, the rankings should become more stable and may still change substantially.
Explore the results
You can use the Image Upscale Arena to compare anonymous pairs on the same 1K-to-4K task and see which results actually hold up. To dig into a specific model or image type, the Image Upscale Compare page lets you pick models and check results against the original.
If you want to check out the current standings with cost, speed and release information, see the leaderboard at /rankings/image-upscale.
Upscale Your Images
Try the top-ranked upscalers yourself in Crafiq studio.