The Problem Is Not Your Upscaler
The most common upscaling complaint is some version of: "I generated a great-looking image, upscaled it, and now it looks worse." The easy assumption is that the upscaler is bad. Usually that is not the problem.
Upscalers do not create detail. They attempt to reconstruct or hallucinate detail that is plausible given the source image. If the source image has soft, muddy, or low-contrast detail, the upscaler is working with bad material. You will get a larger version of a soft image. That is not a bug; that is what upscaling from a low-quality source looks like.
Understanding exactly which failure mode you are dealing with determines the correct fix. Each one has a different cause and a different solution.
Failure Mode 1: Soft Edges and Loss of Fine Detail
What it looks like
Hair, fabric texture, distant foliage, fine architectural detail, text in the background. These elements look sharp at generation resolution. After upscaling, they look like they were slightly blurred. The image is technically larger but perceptually less defined.
What causes it
The source image has less actual sharpness than it appears to have at generation size. At 512x512 or 768x768 pixels, an image has limited total detail. When stretched to 2x or 4x that size, the interpolation fills in the new pixels by averaging adjacent values. If the original detail was at the limit of the generation resolution, it will not survive enlargement cleanly.
Secondary cause: wrong upscaler choice. General-purpose upscalers like Real-ESRGAN standard are trained on photographic datasets. AI-generated images, particularly from Stable Diffusion, have a specific frequency and noise signature that differs from photos. Upscalers trained on AI art will outperform photo-trained upscalers on AI-generated source material.
The fix
First, check your source. Zoom into the original at 100% before upscaling. If it already looks slightly soft at native resolution, the upscaler cannot recover detail that does not exist. You need a sharper source.
To get a sharper source: increase detail in the prompt (add terms like "intricate detail," "sharp focus," "high frequency detail"), increase the step count if you are using a fast sampler, or apply a light sharpening pass (unsharp mask at 0.3–0.5 radius, 0.5–0.8 amount) before upscaling, not after.
For upscaler selection: use Real-ESRGAN x4plus for general use. For anime-style or illustrative AI art, use Real-ESRGAN x4plus-anime. In Topaz Gigapixel AI, the Standard v2 model performs well on AI-generated art; the CG model is also worth testing on high-detail renders. Try two models and compare at 100% crop before committing to a full export.
Failure Mode 2: Plastic Skin or Smeared Smooth Areas
What it looks like
Skin, sky gradients, out-of-focus backgrounds, fabric in broad light. These areas look smooth and natural in the original. After upscaling, they have a slightly plastic, over-smoothed quality. Fine pores or subtle texture are gone. It looks processed.
What causes it
Upscalers trained on high-resolution sharpness targets often interpret smooth gradients as areas to "sharpen" by adding texture or micro-contrast. When the area is correctly smooth (intentional soft light, skin, ambient sky), this added texture reads as artificial. The upscaler is doing its job; it just does not know that this area is supposed to stay smooth.
Over-enhancement strength is the second cause. Most upscalers have a sharpness or detail enhancement slider. Running the strength at maximum does not produce maximum quality. On smooth areas, it produces the plastic effect described above.
The fix
Reduce enhancement strength. In Topaz Gigapixel AI, the Enhancement slider controls how aggressively the model adds detail. For images with significant smooth areas, 40–60% enhancement typically produces cleaner results than 80–100%. The tradeoff is less sharpening on textured areas, which you can partially recover with a targeted sharpening pass in post.
In AUTOMATIC1111 with the Ultimate SD Upscale extension, lowering the denoising strength to 0.2–0.3 (vs. the default 0.4–0.5) significantly reduces the plastic effect on smooth areas while still producing a sharper result than a basic resize.
If only specific areas are affected, Photoshop or Affinity Photo masking lets you apply different enhancement levels to different regions: full strength on detailed areas, reduced strength on smooth areas.
Failure Mode 3: Haloing Around High-Contrast Edges
What it looks like
A faint light outline (a "halo") around dark-on-light or light-on-dark edges: around a figure against a bright sky, around architectural elements, around text. It is subtle at first and becomes more visible when you print or display at full resolution.
What causes it
This is classical oversharpening artifact. The upscaler is applying an excessive local contrast boost at edges, creating a zone of artificially boosted brightness or darkness adjacent to each edge. It is the same artifact that over-sharpened scanned photos have. AI upscalers can produce it when the enhancement strength is too high or when the model has been over-tuned on extremely sharp reference images.
The fix
Lower enhancement strength (same fix as Mode 2 above). Additionally, check if your workflow includes any sharpening step after the upscaler — either in the export settings of your upscaling tool, in Lightroom's output sharpening, or in your image editing software. Applying sharpening twice (once during upscaling, once during export) compounds the haloing. Choose one: sharpen during upscaling at moderate strength, or upscale without sharpening and apply a single targeted sharpen at export. Do not do both.
In Topaz Gigapixel, the "Remove Blur" model is particularly prone to haloing at high strength. The Standard or High Fidelity models produce more controlled edge sharpening.
