The Problem Nobody Talks About With 1920x1080 Fanart
Most people trying to make 1920x1080 fanart end up with something that looks soft and muddy, especially in the shadow areas. They take a low-res source image, run it through an upscaler, and call it done. That is not how it actually works. The difference between something that looks decent and something that looks professional comes down to understanding what information you actually have to start with and what you are asking the tool to fabricate.
A typical source image you find online for a character is maybe 800x450 pixels. It was saved as a JPEG at quality level 60 or lower by someone who uploaded it to Reddit or DeviantArt two years ago. The shadow areas in particular are pure, flat black because the original artist's shading was pushed to maximum compression. You do not have enough data there. Any upscaler will either smooth those shadows out into gray mush or invent edge details that look wrong. This is the core problem, and it determines almost everything you do afterward.
Shadow Fanart Imagesize1920x1080 The Ultimate Guide To High Definition Fan Art
Where to Actually Start
Do not begin with upscaling. Begin with source selection and evaluation. Open your reference image at 100 percent zoom and check the shadow regions specifically. If the shadows are already soft gradients rather than hard blocks of color, you have a better foundation. If they are crushed blacks with no detail, you are going to need to inpaint or redraw shadow information before any upscaling step makes sense. I spent months doing it the wrong way and kept ending up with upscaled images that looked like they had been run through a blur filter in the dark areas.
The working pipeline I use now takes roughly 40 to 70 minutes depending on the complexity of the shadows. Here is what that looks like in practice:
Step one: Open your source in GIMP or Photoshop at native resolution. Create a duplicate layer. Use a soft brush at low opacity to manually reconstruct shadow detail in crushed areas, referencing the original artwork if you have it. This takes 10 to 20 minutes. Step two: Scale the canvas to exactly 1920x1080 using Lanczos resampling, not bicubic or bilinear. Bicubic introduces ringing artifacts around high-contrast edges that then get amplified by the upscaler in the next step. Step three: Run the image through an AI upscaler. The tool choice matters significantly here. I use Topaz Gigapixel AI for photo-realistic fanart and waifu2x-converter-cpp for anime-style pieces. The difference is not marginal. Waifu2x preserves line clarity and handles flat color regions without introducing texture noise. Topaz gives more natural textural detail but can over-sharpen clean anime shading. Both require a dedicated GPU. Running them on integrated graphics is not viable for anything above 1080p output.
Step four: After upscaling, check the shadow areas at 100 percent zoom. You will almost certainly find soft halos or blown highlights where the original shadows were. Select those regions with a refined edge mask and paint in corrected values using a layer set to Multiply blend mode at 15 to 30 percent opacity. Do not use Curves here. Curves shift the entire tonal range and makes the midtones look washed out. Use a Dodge and Burn brush on a separate layer instead. Step five: Export as PNG, not JPEG. A JPEG save at quality 90 still introduces compression artifacts in shadow gradients that become visible when the image is displayed at full 1920x1080 on a monitor. PNG is lossless and the file size difference is usually under 10 MB for this resolution.
Tools and What They Actually Do
Topaz Gigapixel AI uses a model trained on real photography. It excels at restoring detail in skin tones, fabric textures, and environmental elements. It struggles with flat cel-shaded areas and tends to add noise where there should be clean gradients. The default denoise setting of 25 works for most fanart. Anything higher destroys the line art.
waifu2x-converter-cpp is designed specifically for anime and illustration art. It has three processing modes: noise reduction, upscaling, and combined. The combined mode at 2x upscaling with medium noise reduction gives the best results for most 1920x1080 projects. It does not hallucinate new details. It preserves existing lines and flat color regions. The tradeoff is that it produces no texture in shadow areas, which means step one above becomes essential rather than optional.
For people on slower hardware, the free online version of waifu2x works but adds a watermark and caps output at 4x the input resolution. If your source is 640x360, you cannot reach 1920x1080 in a single pass. You need to upscale in stages: 640x360 to 1280x720 first, then 1280x720 to 1920x1080. Skipping the intermediate step creates more artifacts than doing both passes.
Edge Cases That Break Standard Workflows
I encountered a specific problem last year that I have not seen discussed anywhere. A viewer submitted a request for a 1920x1080 piece based on a concept art sheet where the character wore dark armor with intricate embossed patterns in the shadow regions. The source was a 700x394 PNG with some shadow detail preserved. After upscaling, the embossing patterns dissolved into smeared gray shapes. The AI had no reference for what those patterns should look like at higher resolution.
The workaround was to isolate the shadow regions using a color range selection based on luminosity, paste that selection onto a new layer, and then use a small brush to manually trace the pattern details from the original image before running only that layer through the upscaler at 2x. I then composited the upscaled detail layer back onto the fully upscaled base image using a layer mask. This added about 25 minutes to the process but preserved pattern clarity that would otherwise have been lost.
This is not an exceptional case. Any fanart involving detailed shading patterns, chainmail, scales, or ornamentation will exhibit the same behavior. The general rule is: if the shadow area contains repeating geometric or organic detail at the source resolution, upscale that region separately before the full image pass.
Common Pitfalls and What to Avoid
Resizing through Photoshop's Image Size dialog without changing the resample method is the most frequent mistake I see. The default Smart Reconstrain option changes the aspect ratio slightly, and the image ends up at 1920x1018 or similar. Check the dimensions after every resize operation.
Using the Sharpen tool after upscaling instead of before adds harsh edges that conflict with the soft gradient transitions in shadow areas. Apply sharpening selectively with a High Pass filter at 1 to 2 pixel radius, limited to non-shadow regions only.
Ignoring the monitor's actual pixel density. A 1920x1080 image on a 1080p monitor fills the screen exactly. On a 1440p monitor it appears at 75 percent scale and softness issues become less apparent. On a 4K monitor it appears at 50 percent scale and every artifact is visible. Test your output at actual display size, not at 100 percent zoom in your editing software.
Limits of the Method
This approach does not solve a fundamental constraint. If your source image is smaller than 400 pixels on the shortest side, even the best upscaler cannot produce acceptable shadow detail at 1920x1080. The pixel count is too low for the AI model to distinguish between artistic intent and compression noise. At that threshold, the only real option is manual redraw or painting from reference.
AI upscalers also introduce subtle color shifts in dark regions. Shadows that were originally a deep blue-black may shift toward purple or green after upscaling and noise reduction. Always compare the upscaled shadow areas against the original source at the same brightness level. A simple reference window split between the two makes this visible within seconds.
The entire pipeline described above requires approximately 40 to 70 minutes of active work per image at 1920x1080 output. Automated batch processing is possible but sacrifices quality control on individual shadow regions. For a single piece, manual workflow is the only reliable path.
