Photo AI Denoise, Upscale & Restore

The 2022–2026 inflection point: ML-based denoise and upscale beat traditional algorithms by enough that pros switched. DxO and Topaz lead commercial; Real-ESRGAN / Upscayl / waifu2x are the OSS options. Most modern raw developers in Photo RAW Editors now have decent built-in AI denoise too (Lightroom Denoise, darktable's diffuse-or-sharpen, Photomator).

Paid commercial (the gold standard for raw denoise)

• ★ DxO PureRAW — paid perpetual; runs DxO's DeepPRIME XD denoise + lens corrections (see Photo Lens Correction) on raws and outputs DNG. Widely considered the best AI raw denoise in 2024–2026. Bundled into PhotoLab too. Workflow: shoot raw → PureRAW → import DNG into Lightroom / Capture One.
• ★ Topaz Photo AI — paid bundle subscription (or perpetual + upgrade fees). Combines DeNoise AI + Sharpen AI + Gigapixel AI into one app with auto-detection. Strong on edge cases (motion blur, mismatched-noise patterns). Known for occasional over-cooking.
• Topaz Gigapixel AI — paid standalone; the upscaling specialist. Useful when you need 4× / 6× / max output (large prints, archival scans).
• Topaz DeNoise AI — paid standalone; older sibling, mostly subsumed by Photo AI.
• Topaz Sharpen AI — paid standalone; same.
• ON1 NoNoise AI — paid; competent, often discounted.
• Luminar Neo — paid subscription; AI-heavy bundle (denoise, sky replacement, relight, generative). Marketing-loud; results vary.
• Capture One denoise modules — included; Lightroom-class but not Topaz/DxO-class.
• Adobe Lightroom Denoise — included with Lightroom subscription; runs on-device on Apple Silicon / capable GPUs. Not as strong as DxO PureRAW but often good enough.

Free / OSS

• ★ Real-ESRGAN — BSD, Python/PyTorch. The OSS upscale model; pretrained weights for general photos and anime. Often paired with face-restoration models (GFPGAN, CodeFormer). Run locally with the CLI or via Upscayl GUI.
• ★ Upscayl — AGPL, Electron desktop app; wraps Real-ESRGAN + several other models with a one-click UI. Cross-platform, GPU-accelerated. The default OSS upscale GUI in 2026.
• waifu2x / waifu2x-caffe / waifu2x-ncnn-vulkan — MIT; the original anime / line-art upscaler; still useful for low-noise / line-content sources.
• chaiNNer — GPL; node-based image processing GUI; chain Real-ESRGAN + custom workflows.
• GIMP G'MIC — LGPL; massive script library including denoise / upscale filters (some ML-backed via the gmic-py interface).
• Real-CUGAN — anime-focused; outputs are very clean for line work.
• darktable diffuse-or-sharpen, denoise (profiled), denoise (non-local means) — included in darktable; strong traditional denoise.
• RawTherapee Noise Reduction — included; long-standing.
• Lensfun + ImageMagick — for sharpen / lens-corrected output via classic algorithms.

Specialty

• GFPGAN / CodeFormer — face restoration models; restore badly compressed or blurred faces. Often combined with Real-ESRGAN.
• CodeFormer-WebUI — local GUI for the above.
• NafNet, MIRNet, Restormer — research-grade restoration models; mostly Python notebooks.
• Pixelmator Pro Super Resolution — paid macOS; ML upscale, well-integrated.
• Photomator Super Resolution — same idea, Apple-owned.

Hosted / API

• Replicate — pay-per-call hosted Real-ESRGAN, GFPGAN, CodeFormer, custom restoration models. See AI / LLM for the broader hosted-AI ecosystem.
• Fal.ai — same idea; faster cold-start.
• remove.bg API — single-purpose; closed, paid.
• Topaz hosted API — paid; for high-volume pipelines.

Workflow patterns

• Denoise the raw, not the JPEG — DxO PureRAW outputs a clean DNG that Lightroom / Capture One then develop normally. Always denoise before tone-mapping / sharpening.
• Upscale only on output — never upscale the master; upscale a copy at export size.
• Match noise across composites — when you're stitching frames or compositing, denoise before the merge so noise patterns aren't mixed.
• Expect occasional artifacts — AI denoise sometimes invents detail (smooth eyes, plasticky skin, fake leaves). Inspect at 100%.
• GPU matters — Real-ESRGAN and Topaz are GPU-bound; an RTX 3060 / Apple Silicon M1+ is a meaningful threshold.

What's changing in 2024–2026

• On-device AI denoise on Apple Silicon — Lightroom, Photomator, Pixelmator Pro now run high-quality neural denoise locally without cloud round-trips. M3 / M4 Pro / Max are the sweet spot.
• GenAI-driven inpainting / generative remove — Adobe Generative Fill / Generative Remove in Lightroom; Photoshop content-aware. Still cloud-based for Adobe; on-device experimental.
• Topaz "Photo AI" generative features — beyond denoise / upscale; selective subject-aware corrections.
• OSS upscalers improving fast — Real-ESRGAN+ and 2024 community-trained models close the gap with Topaz on most natural photos; Topaz still wins on truly difficult sources.
• Tooling fragmentation pressure — Lightroom + DxO PureRAW + Topaz Photo AI + Photoshop is a 4-app workflow many pros consolidate to 2.

Pick this if…

• Best raw denoise, no subscription: DxO PureRAW (perpetual).
• Bundle of denoise + sharpen + upscale, paid: Topaz Photo AI.
• Free, runs locally, GUI: Upscayl (Real-ESRGAN under the hood).
• Free, scriptable / pipeline: Real-ESRGAN CLI directly.
• Already in Lightroom subscription: Lightroom Denoise is "good enough" for most.
• Face restoration on old / scanned photos: GFPGAN or CodeFormer.
• Anime / line-art upscale: waifu2x or Real-CUGAN.
• Hosted in a SaaS pipeline: Replicate or Fal.ai.