The State of Play

AI-driven upscaling, restoration, and compositing has consolidated at leading-edge status after six years at this tier — technically functional but deployment remains bounded by unresolved hallucination and fidelity barriers. The practice demonstrates mature ecosystem with vendor commoditization, broad tool accessibility, and documented production success in tolerance-forgiving domains (e-commerce, media remastering, asset restoration). Simultaneously, practitioner skepticism persists unchanged: restoration models generate statistically plausible details rather than recover originals; upscaling actively hallucinate missing content; identity drift and ethnic-feature bias remain systematic failures in heritage applications. The practice works at scale where output quality variance is acceptable (product photos, thumbnail remastering), but authenticity and accuracy concerns prevent advancement into general-purpose professional workflows requiring pixel-level fidelity. Vendor consolidation (Adobe-Topaz integration, AWS Bedrock entry, subscription-only licensing) has not resolved the technical-limitation-to-advancement barrier that has blocked this practice from reaching good-practice tier for six consecutive years.

Ecosystem maturation has accelerated via vendor consolidation and API-first commercialization through June 2026. Adobe Photoshop 27.8 (June 2026) ships Firefly Upscale bug fixes and new image models (Firefly Image 5, Flux.2 Pro, Gemini), refining composite workflows alongside Topaz Labs partnership integration from v27.0. AWS Bedrock continues Stability AI upscaling with three variants (Fast $0.02, Conservative $0.40, Creative $0.60). Topaz Labs shipped largest-ever model release (six models: Wonder 3, Denoise Max, High Fidelity 3, Super Focus 3 in April 2026) with June releases (Photo 1.6.1, Gigapixel 1.3.1) delivering 4–4.4× speed improvements via NeuroStream acceleration—6MP→24MP upscaling now achievable in 7.5 minutes on RTX 4060 laptop (vs 33 minutes previously). Topaz Image Upscale API (released April 2025, widely distributed via Replicate, fal.ai, Lumenfall) demonstrates multi-platform commoditization: five specialized models with tiered pricing ($0.05–$0.82 per output MP), ranked #3 in community leaderboard (Crystal Upscaler 1278 Elo, Recraft 1173, Topaz 1125) by blind voting (61+ battles). Platform consolidation deepens: Replicate curates multiple vendor upscalers across production use cases; Photta platform reports 24k+ sellers generating 95k+ images in 3 months with 43% weekly growth. E-commerce dominance confirmed: 62% of consumers identify fully AI-generated lifestyle photos; hybrid compositing (real hero shot + AI lifestyle staging) achieves 3.1× higher conversion—validating composite workflows as production strategy. Production adoption benchmarks: Netflix 38% of catalog (1,200→72 hours/film), Warner Bros documentaries, Decathlon (99% cost reduction), Mercari (1% uplift at 10% adoption). Yet fundamental limitations remain unresolved. Adobe Generative Upscale produces hallucinations (maps with invented rivers, incorrect shapes). ON1 Restore AI exhibits face repainting and color reimagining without faithful preservation. Fine-art printing expertise documents upscaling as exposure risk rather than improvement path: native 200–300 ppi preferred over interpolated, signaling domain-specific limitations. Practitioner displacement emerging: Topaz denoising dominance eroding to DxO PureRAW and Lightroom AI Denoise. Independent testing across 300+ photos: "statistically probable guesses—not factual reconstructions." Practitioner assessment: 62% of visualization professionals report tools not fully production-ready, 77% cite inconsistency. Identity drift, genetic alteration bias, and missing-region hallucination persist in heritage applications. Lightroom Super Resolution incompatible with Denoise due to operation-chaining limits. Research advances (CVPR 2026) focus on efficiency gains (16× GPU memory reduction) and perceptual quality metrics via human-preference modeling, yet do not resolve fidelity barriers. The practice demonstrates measurable ROI in tolerance-forgiving domains (e-commerce, restoration workflows), yet authenticity concerns and unresolved hallucination risk prevent advancement beyond bounded use cases despite six consecutive years at leading-edge tier.

• 2019: Deep learning for image super-resolution and restoration published comprehensive surveys (IEEE TPAMI), organized benchmark competitions (AIM 2019 Challenge), and demonstrated cross-domain applications in satellite imagery. Fundamental instability issues in deep learning image reconstruction were documented. Adobe launched AI-powered upscaling in Lightroom; Topaz Gigapixel AI was commercially available but computationally expensive and slow. Industry practitioners acknowledged AI limitations compared to human restoration expertise, indicating unresolved challenges in image quality and artistic reconstruction.

• 2020: Production deployments expanded significantly—Pixar deployed GANs for learned resolution, reducing rendering time from 50 CPU hours to 15 seconds per frame and cutting render-farm footprint by 75%; Facebook deployed neural super-sampling for real-time VR upscaling to 2K. Microsoft released two CVPR 2020 papers and open-source implementations for reference-based super-resolution and old photo restoration with face enhancement. Commercial adoption by professional photographers broadened (Topaz Gigapixel AI in professional workflows), and service-based offerings emerged (Neural Love for historical media restoration). However, critical limitations persisted: processing remained slow (minutes to hours), quality was input-dependent, and ethical concerns arose around historical media authenticity. The practice demonstrated viable production viability in media/entertainment yet faced adoption barriers from computational cost and reliability concerns.

• 2021: Research community advanced restoration architectures (Restormer Transformer achieving SOTA across 16 tasks) and explored generative approaches (StyleGAN2-based Time-Travel Rephotography unifying restoration workflows). Google published SR3 diffusion models with 50% human confusion on 8x upscaling, advancing photorealism metrics. Mainstream adoption accelerated: Adobe integrated Super Resolution into Lightroom Classic, and competitive market ecosystem solidified (Photoshop, Pixelmator Pro, Gigapixel AI). However, fundamental limitations remained unresolved—colorization lacked historical context, inference speed remained practical barrier despite improvements (Gigapixel 5.5.0 353% faster but still minutes for large files), and ethical concerns around authenticity persisted for historical media applications.

• 2022-H1: Vision Transformers emerged as preferred restoration architecture across 7 tasks (super-resolution, denoising, enhancement, artifact reduction, deblurring, adverse weather, dehazing) per comprehensive surveys in Sensors and Neurocomputing journals. CVPR 2022 NTIRE workshop demonstrated continued innovation momentum with efficiency challenges. Ecosystem expanded with enterprise cloud platforms (Viesus Cloud reporting 58% complaint reduction at Albelli, 15-second processing at Swiss-Image). Topaz Gigapixel AI v5.8 released with GPU acceleration and memory improvements. However, professional adoption hesitation persisted: photographers questioned Super Resolution quality for stock submissions, and TechCrunch testing revealed face artifacts and unrealistic sharpening in Picsart's enhancer, indicating unresolved output quality concerns limiting commercial deployment.

• 2022-H2: Product ecosystem matured with specialized tooling: Let's Enhance launched Smart Resize for e-commerce (6x upscaling with text preservation), while independent practitioner case studies showed successful deployments (bird photography at 6000x4000 to 12000x8000, vintage photo restoration). However, critical barriers remained: Adobe Super Resolution restricted to RAW files with massive output files (182.9MB for 19.6MB source), chromatic aberration issues; Topaz Photo AI demonstrated quality trade-offs with overaggressive face reconstruction ("uncanny valley" artifacts); competitive tools (Luminar NEO Upscale AI beta) showed significant performance gaps (3x slower, inferior sharpness). Hardware benchmarking confirmed computational intensity of commercial tools. Practical deployment continued to expand despite unresolved quality-consistency and performance limitations.

• 2023-H1: Research advanced toward multi-task restoration architectures (DaAIR framework) capable of handling multiple degradations simultaneously. Adobe expanded Enhance suite with AI-powered Denoise feature (April 2023) using deep CNN optimized for NVIDIA TensorCores and Apple Neural Engine. Topaz Photo AI released v1.2 with architectural improvements and larger model size. Market research confirmed adoption growth in e-commerce, social media, and digital marketing with North America leading, Asia-Pacific expanding. However, professional adoption barriers remained entrenched: practitioner forums revealed persistent tool trade-offs (variable performance across image types), Wikimedia Commons policy debate reflected broader authenticity concerns about AI-enhanced content, and professional skepticism continued regarding quality reliability for commercial workflows despite vendor product advances.

• 2023-H2: Ecosystem continued maturation with specialized vendors (Puget Systems) benchmarking professional tool performance on enterprise-grade hardware (NVIDIA RTX 6000 Ada), signaling hardware optimization for production workflows. Research advanced facial restoration with DAEFR framework addressing perceptual-distortion trade-off through dual-branch architecture. However, critical adoption barriers persisted and intensified: ethical concerns about historical misuse escalated (former editor documented ease of AI-driven alterations to iconic photos), practitioner testing revealed print-quality limitations (Adobe Super Resolution enables enlargement but not perceptual sharpness improvement), and real-world deployment in e-commerce faced quality assurance challenges (stock platform rejections due to upscaling artifacts and noise in AI-generated content). Archival specialists raised concerns about AI's role as "remixing" rather than restoration, citing identity shift and ethnic feature bias in face restoration. The window closed with the practice demonstrating viable technical capability in specific workflows but fundamental maturity barriers in authenticity, bias mitigation, and quality reliability remaining unresolved across professional deployment contexts.

• 2024-Q1: Ecosystem expanded with new cloud-based tools (Media.io 8x upscaler, Kittl integrated upscaler to 4096x4096) and market research confirming sustained growth across e-commerce and digital marketing verticals with Asia-Pacific region accelerating. Major vendors deepened professional adoption support: Adobe's Super Resolution workflow integration into Lightroom Classic demonstrated sustained professional investment, enabling real-world deployments (3.1MP to 12.4MP scaling in DAM pipelines). However, user feedback revealed persistent tool limitations: Topaz community forums documented unresolved fidelity concerns with existing upscaling models, requests for scale limits beyond 6x, and feature inconsistency criticism. Critical analysis maintained skepticism about restoration boundaries: technical assessment documented fundamental impossibility of recovering irreversibly lost pixel data without original sources, reinforcing that AI restoration remains approximation-driven. The quarter demonstrated continued ecosystem growth and vendor capability expansion, yet unresolved quality consistency and technical restoration limits persisted as adoption barriers in professional workflows.

• 2024-Q2: Ecosystem maturation continued with architectural innovation: Topaz Gigapixel AI 7.1.0 (April 2024) introduced diffusion-based Recovery model specifically designed for low-resolution image upscaling, expanding model diversity beyond traditional CNN/Transformer approaches. Consumer-facing adoption expanded with FixPhotos.ai demonstrating service-scale deployment (252k+ photos restored, 30k+ customers). However, critical barriers persisted: Adobe Photoshop beta testing revealed unresolved quality artifacts in photo restoration (t-shaped distortions, quadrant division), while Lightroom Classic users reported technical compatibility issues with Super Resolution feature (CR3 format graying out). The quarter showed continued vendor innovation and service-based adoption growth alongside persistent quality-consistency and technical-integration challenges limiting broader professional deployment.

• 2024-Q3: Product ecosystem continued refinement with performance focus: Topaz Gigapixel AI 7.2.0 and 7.3 released with 20x+ faster Recovery mode, expanded model options (8 models including High Fidelity, Art & CG, Recovery variants), CMYK support for print workflows, CLI access, and new commercial Pro licensing ($499/year). Pixa (rebranded Pixelcut) launched commercial Image Upscaler API with $0.1-per-image pricing, indicating API-first commercialization trend. However, real-world deployment barriers remained persistent and critical: Adobe's Super Resolution and Denoise features documented significant workflow integration issues (orphaned cache files, 10+ minute processing times, 95% CPU usage, disruption to masking workflows), and educational guides emphasized that despite tool accessibility, AI restoration requires significant skill and manual adjustment to achieve quality results, with inherent limitations (blurriness, color inaccuracy, artifacting) remaining unresolved. The quarter demonstrated accelerating vendor feature development and commercialization velocity alongside unchanged fundamental quality-consistency and workflow-integration barriers limiting mainstream professional adoption.

• 2024-Q4: Ecosystem maturation continued with evidence of enterprise-scale deployment: market research documented production adoption across major media companies (Netflix upscaling 38% of catalog, reducing restoration time from 1,200 to 72 hours per film; Warner Bros remastered 1999 documentary to 4K) and e-commerce platforms (Alibaba achieving 17% conversion uplift after AI enhancement). Research community advanced applications with peer-reviewed studies on photogrammetric integration, while commercial expansion continued (Upscale.media 8x web upscaler, continued Topaz model proliferation). However, critical adoption barriers persisted unchanged: software reliability issues emerged (Gigapixel v8.0.3 GPU failures with ONNX errors), professional workflow integration remained limited (Adobe feature reliability issues ongoing), and authenticity concerns continued to constrain professional adoption despite demonstrated capability. The quarter closed with the practice demonstrating viable deployment at scale in media restoration and growing adoption in e-commerce and photography verticals, yet persistent reliability, workflow integration, and authenticity barriers remained unresolved at year-end 2024.

• 2025-Q1: Ecosystem expanded with vendor proliferation across specialized niches (Magnific for AI-generated art, Upscayl free/open-source, Topaz photo-focused, Lummi Ultra high-resolution to 8,600px, Freepik web-based, HitPaw beginner-friendly). Market research signaled sustained growth with AI photo restoration/colorization market projected to reach $5B by 2027 (25%+ CAGR). However, professional adoption barriers remained entrenched unchanged: Deloitte TMT analysis documented that major studios remain cautious about deploying image/video AI for production due to tool immaturity, IP liability, and defensibility concerns despite technical capability maturity. Practitioner assessment confirmed that despite widespread tool availability, AI upscaling remains an educated guess producing unnatural textures, plastic artifacts, and tiling issues requiring manual correction. Adobe's major tool (Lightroom Super Resolution) clarified scope limitation: enlarges resolution only without improving quality, remaining ineffective for noise reduction, blur correction, or detail recovery. The quarter demonstrated continued ecosystem expansion and market confidence in adoption growth, yet fundamental professional deployment barriers—reliability, workflow integration, authenticity concerns, and inherent quality limitations—persisted unchanged, with major production adopters (studios, platforms) deferring full production adoption pending maturity advancement.

• 2025-Q2: Research community advanced creative upscaling with C-Upscale diffusion method for ultra-high-resolution generation (8,192×8,192) using global-regional priors, signaling continued academic innovation. Commercial ecosystem continued feature evolution: Topaz Gigapixel v8.4.0 split Redefine model into Realistic and Creative modes with personalized learning, extending vendor product maturity. Market quantification confirmed sustained growth momentum: AI image processing market at USD 2.42 billion in 2025, 10.53% CAGR to USD 4.88 billion by 2032, signaling broad adoption across consumer electronics, automotive, medical, and security verticals. However, critical limitations persisted and deepened perception challenges: independent technical evaluation (Furnets) documented systematic issues across deployed models—plastic texture hallucination, edge overwrites, extreme processing slowness—with real-world images, reinforcing that commercial tools cannot reliably reconstruct natural detail. Historical photograph restoration failures (ChatGPT attempted restoration of 1826 photograph) demonstrated continued hallucination and accuracy risks on culturally significant material. Major ecosystem gap emerged: Google discontinued Imagen 1/2 upscaling support, signaling deprioritization by major vendor despite practice importance. The quarter demonstrated research advancement and continued vendor feature development alongside unresolved technical limitations and shifted ecosystem focus limiting broader professional adoption.

• 2025-Q4: Vendor ecosystem maturity advanced with strategic partnership: Adobe Photoshop integrated Topaz Labs models (Gigapixel and Bloom) as native Generative Upscale feature, signaling ecosystem consolidation and mainstream adoption pathway. Market forecasts remained growth-oriented: super-resolution market projected $6.72 billion by 2033 (18.9% CAGR) with North America at 38% market share and Asia-Pacific fastest growth at 21.5% CAGR. However, critical deployment barriers persisted unresolved at year-end: independent practitioner testing documented persistent limitations (upscaling "only works with good focus" on small enlargements, larger increases "look really fake" with artifacts), and real-world deployment failures emerged (Gigapixel AI user reports purchasing yearly subscription then experiencing "terrible deception" with minimal improvement and non-functional Face Recovery). Tool ecosystem remained specialized with trade-offs (Topaz Photo AI all-in-one to 4x; Gigapixel specialized at 6x detail preservation) rather than unified excellence. The quarter demonstrated continued vendor investment and strategic partnerships alongside unresolved technical and workflow reliability barriers limiting mainstream professional adoption despite five years of category maturity.

• 2026-Jan: Vendor partnerships faced execution challenges as Adobe's Generative Upscale remained available only in Photoshop beta (not stable release 26.11), indicating incomplete production rollout despite 2025 announcements. Critical reliability issues emerged: Adobe's Generative Upscale produced hallucinations on accuracy-critical content (maps with invented rivers and incorrect mountain shapes), demonstrating fundamental limitations on specialized use cases. Topaz Labs shifted business model, abandoning perpetual licenses for Gigapixel AI in favor of subscription-only ($50-69/month), generating user dissatisfaction and highlighting pricing barriers. Ecosystem analysis confirmed continued adoption friction: 62% of visualization professionals report AI not fully production-ready, 77% cite inconsistency as major concern, while broader creative adoption remained concentrated in e-commerce and entertainment (Netflix, Warner Bros) rather than mainstream professional workflows. The month demonstrated ecosystem maturity in tooling alongside persistent execution, reliability, and cost barriers limiting broader production adoption.

• 2026-Feb: Market expansion accelerated with projections for AI image enhancement tools reaching $88.7B by 2025 and $50.7B in the broader image enhancer market by 2034 (34.6% CAGR), signaling strong investor confidence. However, critical deployment barriers persisted and sharpened: Adobe's integration of Topaz models into Photoshop remained feature-limited, workflow constraints worsened (Super Resolution and Denoise features incompatible due to operation chaining limits), and critical evaluations testing 14 tools across 300+ photos documented that AI restoration produces "statistically probable guesses—not factual reconstructions" requiring hybrid human-AI workflows. Topaz released refined Gigapixel AI with improved Face Recovery models (Creative/Realistic variants) and CLI support, yet market transition to subscription-only model created adoption friction. The month demonstrated market-level confidence in category growth alongside persistent professional skepticism about reliability and quality consistency, with fundamental technical limitations preventing general-purpose deployment.

• 2026-Mar: Production adoption continued expanding at use-case-specific scale. Fine art photography (VanSky Studio) integrated Topaz Photo AI for high-ISO recovery enabling 87% acceptance rate on previously marginal frames, reducing culling from 8 hours to 40 minutes. E-commerce documented transformational ROI: Maya Chen (sustainable fashion) achieved 94% cost reduction ($28→$1.85 per image) scaling from 150–350 to 2,847 SKUs/month through AI enhancement workflows. Photography industry adoption reached 90% for post-processing automation and 74% for AI noise reduction (PhotoWorkout survey). Heritage restoration deployments expanded: University of Rome La Sapienza deployed AI for Colosseum structural analysis and artifact preservation. Topaz Labs released API with five new models (Starlight variants, Background Removal, Gaia 2) and unified pricing, signaling ecosystem maturity and broader model accessibility. However, critical limitations persisted at highest professional standards: York University documented systematic failures on historical photograph restoration—hallucinated elements, anachronistic details, ethnic feature bias in face recovery. Adobe's practitioner base reported deepening adoption friction: vendor shift to AI-first roadmap with metered upscale pricing generating dissatisfaction among legacy users. The month closed with the practice demonstrating viable, high-ROI deployment in tolerance-forgiving domains (e-commerce, asset restoration) and measurable efficiency gains in professional creative workflows, yet persistent authenticity risks and adoption friction for highest-fidelity applications.

• 2026-Apr: Vendor ecosystem consolidation accelerated with major platform integrations, new enterprise entrants, and Topaz Labs' largest-ever model release — six new models (Wonder 3, Denoise Max, Super Focus 3, High Fidelity 3) shipped simultaneously in the Next-Gen launch. AWS launched Stability AI Image Services on Bedrock (April 2026) offering three upscaling variants (Fast $0.02/img, Conservative $0.40/img, Creative $0.60/img), signaling Tier-1 cloud vendor commitment to production-ready upscaling. Adobe Photoshop v27.0 moved Generative Upscale from beta to native integration of Topaz Gigapixel & Bloom, enabling 4x upscaling to 56MP+ with detail retention and Harmonize feature for automated compositing color-matching. Photoroom Intelligence launched globally with documented case studies: Decathlon achieved 99% cost reduction and week-to-minutes processing; Mercari reported 1% listing uplift at 10% seller adoption. Research ecosystem remained robust: NTIRE 2026 (CVPR 2026 workshop) attracted 100+ teams and 3,000+ submissions for low-light portrait restoration; the LoViF 2026 Challenge drew 124 participants advancing unified all-in-one restoration models. However, critical fidelity barriers persisted unresolved: ON1's new Restore AI tool exhibited systematic hallucinations in face restoration and color reimagining rather than faithful preservation; independent analysis confirmed AI upscalers actively generate plausible details via learned inference rather than recovering lost information, reinforcing the identity-drift and hallucination risk that caps the practice at leading-edge despite consolidating ecosystem and expanded enterprise accessibility.

• 2026-May (14–28): Vendor ecosystem reached inflection point with major platform integrations and practitioner testing validating limited scope of viable deployment. Adobe Photoshop 27.7 (May 26) moved Generative Upscale from beta to native feature via Topaz Labs partnership and launched on-device Remove tool (~5GB AI model) enabling local object removal without cloud processing, signaling shift toward local inference. Topaz Labs released Photoshop plugin GA for Gigapixel with 30,000px max upscaling and face recovery; simultaneous release of Topaz Photo 1.6.0 and Gigapixel 1.3.0 with NeuroStream 2 acceleration (2–4x speedup on diffusion models) and Noise-Aware Sharpening model separating noise from recoverable detail. Professional practitioner testing confirms ecosystem maturity: Finding the Universe (May 2026) documented that image-quality gap between Topaz Photo, DxO PureRAW 6, and Adobe Lightroom 15.3 has closed to pixel-level inspection on high-ISO files, indicating commoditization of denoising capability. Expert reviews (lartdelaphoto.fr) position Topaz Photo as 2026 reference standard; L'Art de la Photo Gigapixel review confirms High Fidelity model for faithful restoration, Recover model for degraded files, and Redefine for generative detail—practical guidance emphasizing model selection per use case. Topaz Image Upscale API deployment (5 specialized models, 600% scaling, $0.025/invocation) demonstrates ecosystem maturity through commercial API distribution; Replicate platform curation of multiple upscaler vendors signals multi-vendor production adoption. Critical assessment emerged: Artedge AI FAQ documents realistic limitations (4× upscaling risky, severe motion blur rarely fixable, waxy skin and identity drift as persistent failure modes) balancing vendor optimism with unresolved practitioner experience of hallucination and artifact risks. NAB 2026 conference discussion (Topaz Labs exec) documented real-world deployments on Babylon 5 upscaling and documentary restoration with Apple Silicon optimization, validating production adoption in high-visibility projects. Cost-benefit analysis (BigImg.ai) established that Real-ESRGAN free option achieves 90% quality vs. Topaz Gigapixel ($99 commercial), shifting adoption criteria from tool access to implementation strategy and use-case specialization. The May 14–28 window closed with evidence of commoditized tool ecosystem and specialized deployment success, yet practitioner skepticism about general-purpose quality reliability persisted despite vendor feature velocity. Note: Practice remains at leading-edge despite ecosystem maturation—continued platform integration and specialized model proliferation have not resolved fundamental hallucination/fidelity trade-offs preventing advancement to good-practice tier.

• 2026-Jun: E-commerce adoption hit first-party scale benchmarks: Photta platform reports 24k+ sellers generating 95k+ images in 3 months at 43% weekly growth, and industry survey data shows 67% of leading e-commerce operators budget for AI imaging with 87% reporting revenue uplift — confirming production-scale deployment in the tolerance-forgiving segment. Consumer detection of fully AI-generated product photos reached 62% (up from 38% in 2023), while hybrid compositing (real hero shot + AI lifestyle staging) achieves 3.1x higher conversion, quantifying the production ROI boundary. Adobe Photoshop 27.8 (June 2026) ships Firefly Upscale bug fixes and adds Firefly Image 5, Flux.2 Pro, and Gemini model options; Topaz released Gigapixel 1.3.1 (4.4x speedup on RTX 4060 via NeuroServer) and Photo 1.6.1 (NeuroStream 2, 2-4x diffusion speedup), with the Topaz Image Upscale API now distributed across three cloud platforms (Replicate, fal.ai, Lumenfall) demonstrating API commoditization. CVPR 2026 "Upsample Anything" (KAIST/MIT/Microsoft) introduces training-free upsampling reducing GPU memory 16x while maintaining quality, advancing on-device deployment. Community benchmarking (Lumenfall Arena, 61+ blind battles) places Crystal Upscaler first (1278 Elo), Recraft second, Topaz third — alongside a practitioner review documenting Topaz noise reduction dominance eroding to DxO PureRAW and Lightroom AI Denoise. Ecosystem analysis confirms upscaling as commodity, while CVPR 2026 research continues to identify hallucination and structural fidelity trade-offs as unresolved core barriers (post-processing fusion required for pixel-level consistency), reinforcing that the fidelity ceiling preventing good-practice advancement remains unchanged despite commoditized tooling.