The State of Play

AI-powered inpainting and outpainting have achieved commercial production scale without solving the fundamental reliability ceiling. July 2026 data confirms mainstream adoption: Firefly reached $300M ARR with 50% quarter-over-quarter growth; 75% of 16,000 surveyed creators report AI as integrated or essential to workflow; agentic orchestration (Firefly AI Assistant) now automates multi-step inpainting workflows across Creative Cloud. Yet production reality exposes persistent constraints: 57% of creators report outputs require moderate-to-extensive editing before publication; retailers (Target, H&M, Amazon) require manual correction on 60-80% of inpainting outputs due to edge artifacts; professional photographers retain 2.4-3.1× higher conversion rates than AI-assisted alternatives. The practice stays leading-edge because feature ubiquity and commercial viability have decoupled from production reliability—systems work reliably for 80% of cases (backgrounds, object removal in optimal conditions) but fail predictably on edge cases (complex geometry, fine details, identity preservation), forcing hybrid human-AI workflows with mandatory quality-control gates.

July 2026 confirms the practice has achieved commercial scaling without resolving production reliability gaps. Firefly ARR hit $300M in Q2 2026 with 50% QoQ growth; Firefly AI Assistant (GA June 18) orchestrates multi-step inpainting/outpainting workflows across Photoshop, Premiere, Illustrator, InDesign with agentic multi-app automation. Independent evaluations affirm production maturity: Ridge Marketing's three-year retrospective documents Firefly's evolution from unreliable to professional-grade output; independent hands-on testing (PAGE Online, ad-hoc-news) confirms 2K resolution, fewer artifacts, improved edge consistency. Creator adoption is substantial and confirmed: Adobe's Creators' Toolkit Report (n=16,000) shows 75% of creators describe AI as integrated/essential; 84% of photographers use AI for time-saving on routine tasks; Pixelcut reached 2M+ users, Getimg.ai 10M+ users. FAL.ai ecosystem reports 985 model endpoints with FLUX.1 Kontext and Qwen Image Edit providing production-quality text-guided inpainting. Enterprise governance frameworks are maturing: third-party compliance assessments confirm Flux deployment meets GDPR requirements with EU residency options, SOC 2 unknown, enabling enterprise adoption.

Yet fundamental quality and reliability constraints persist and block tier advancement. Production deployment metrics document the ceiling: 57% of creators report AI outputs require moderate-to-extensive editing before publication (Adobe Creators' Toolkit Report, June 2026); retailers (Target, H&M, Amazon) continue requiring manual correction on 60-80% of inpainting outputs due to edge smearing and artifact hallucination (Rewarx analysis, June 2026); professionally composed photos outperform AI-assisted versions by 2.4-3.1× (Rewarx, June 2026), with generic AI backgrounds reducing conversion by 28%. Real production deployments document specific failure modes: LaMa video localization pipeline achieved SSIM 0.37 (starting from 0.15), targeting 0.85 with metric-driven iteration; professional retouching confirms reliable automation (culling, color correction, background isolation) but persistent failures (skin texture, hand anatomy, jewelry fidelity, ghost mannequin compositing). Platform reliability gaps emerged: persistent service failures in Firefly GA despite paid credits (Adobe Community, June 2026); design tension between full-frame content recreation and pixel-level fidelity preservation constrains professional adoption (expert critical assessment, June 2026). Negative signal remains structural: PaintBench benchmark documents 17.1% mIoU on precise visual editing tasks (83% failure rate); MediaEval 2025 recognizes inpainting as major manipulated-image vector with 75,000 synthetic test images. Social media adoption metrics reveal authenticity backlash: 340% increase in user-flagged 'not real' comments following advanced texture-removal tools; 22% decline in save rates for AI-smoothed content; algorithmic penalties for over-polished faces in recommendation feeds (Remio analysis, July 2026). The practice demonstrates commercial viability (Firefly $300M ARR, 75% creator adoption, agentic orchestration) and research sophistication (efficiency breakthroughs, preference alignment, forensic detection) while reliability constraints at edge cases—complex geometry, fine details, identity consistency—force hybrid human-AI workflows with mandatory quality-control gates, preventing autonomous production-scale deployment needed for tier advancement.

• 2022-H2: DALL-E 2 and Stable Diffusion introduced dedicated inpainting and outpainting features, enabling AI-driven image editing and canvas extension. Early tutorials and community experimentation began documenting these capabilities, marking the foundation of the practice.

• 2023-H1: Major vendors entered the market and production barriers became visible. Adobe shipped Generative Fill in Photoshop beta (May 2023), while AWS enabled enterprise Stable Diffusion inpainting deployment via SageMaker JumpStart (June 2023). Academic research on inpainting security defenses indicated technology maturity, but widespread bug reports from both open-source and commercial platforms revealed reliability gaps: outpainting producing artifacts, inpainting failing on non-contextual edits, and content filtering limiting workflows. Feature availability accelerated while production-grade stability lagged.

• 2023-H2: Feature proliferation and mainstream recognition accelerated despite ongoing quality issues. Adobe's Generative Expand reached GA and earned TIME's Best Invention recognition, with adoption scaling to 3+ billion images generated. OpenAI launched DALL-E 3 with improved inpainting, while research publications (ICML, ACM MM) advanced technical capabilities on coherence, multimodality, and resolution. However, production reliability remained constrained: practitioners documented systematic quality issues including resolution limits (1024x1024), unwanted object insertion, color inconsistency, and restrictive content filters. The gap between widespread feature availability and production-grade quality defined this period.

• 2024-Q1: Platform integration accelerated while reliability concerns surfaced. OpenAI expanded inpainting/outpainting to ChatGPT (March 2024), integrating the practice into a mainstream conversational AI platform. E-commerce applications advanced with Amazon publishing real-time virtual try-on research targeting billions-of-products scale. Technical research addressed performance barriers: 4K-resolution algorithms achieving >60 fps, and mask-aware diffusion improvements (Soft Inpainting) in open-source tools. However, quality regressions emerged: Photoshop users reported systematic failures in subject removal and speculation about training data rollbacks due to consent concerns. The quarter demonstrated continued feature proliferation offset by instability in production deployments, suggesting the practice remains in the painful phase where availability outpaces reliability.

• 2024-Q2: Vendor iteration and ecosystem maturation accelerated alongside persistent reliability challenges. Adobe released Firefly Image 3 (April 2024) with improved quality and reference-image guidance. Hugging Face standardized outpainting documentation in diffusers library (May 2024). Academic research advanced bias mitigation and transformer-based techniques (CVPRW, arXiv). However, production deployment barriers persisted and expanded: Photoshop users reported continued "Resource not available" connectivity failures, while SDXL inpainting exhibited new noise artifacts at maximum strength. By quarter-end, the practice showed feature abundance with unresolved control and reliability gaps, constraining confident production deployment.

• 2024-Q3: Research and security focus intensified as maturity concerns surfaced in academic literature. CVPR 2024 showcased instruction-guided editing workflows combining inpainting with language models, demonstrating emerging sophistication in application design. Peer-reviewed research at IJCAI and arXiv advanced understanding of inpainting vulnerabilities: adversarial attacks against Stable Diffusion inpainters, forensic detection methods for identifying tampered regions, and weaknesses in proposed defenses. Meanwhile, capability gaps persisted in enterprise offerings—Azure OpenAI confirmed non-support for DALL-E image editing APIs through mid-2024. The quarter demonstrates a pattern of widening technical sophistication coupled with persistent reliability and deployment challenges: academics advancing attack surfaces and detection, while practitioners encountered continued quality regressions and missing enterprise integrations. The practice remained in the "uneven maturity" phase—research depth accelerating while production confidence stagnated.

• 2024-Q4: Vendor feature announcements coincided with critical reliability regression. Adobe announced updated Generative Fill and Expand features at MAX 2024 (October), prompting expectations of improved quality; instead, professional users reported dramatic degradation, with success rates dropping from 90-95% to 5-10% post-update, accompanied by crashes and feature failures. Capability gaps persisted: Azure OpenAI still lacked inpainting support (confirmed October). Applied research expanded into domain-specific use cases (cultural heritage restoration), and open-source community development standardized complex multi-step workflows (SDXL outpainting with upscaling via ComfyUI). The quarter exemplified the practice's core maturity gap: sophisticated research and developer tooling advanced while the primary commercial platform regressed in production stability, reinforcing that feature abundance and deployment reliability remain decoupled.

• 2025-Q1: Academic research matured significantly while production reliability remained constrained. IEEE TPAMI published a comprehensive diffusion-model survey consolidating inpainting/outpainting methodologies; CVPR and ICCV 2025 papers advanced semantic correctness (SAGI dataset of 95k+ images), object addition via inverted inpainting, and forensic detection. Applied deployment expanded into resource-constrained environments with quantization optimization for edge inference on agricultural data augmentation. However, Adobe Generative Fill continued to exhibit quality issues: users reported overzealous content filtering, unwanted object insertion, and degradation compared to beta releases throughout Q1. Enterprise cloud integration remained incomplete: Azure OpenAI still lacked inpainting support. The quarter widened the gap between academic sophistication and production stability—theoretical foundations solidified while commercial reliability and content filtering remained barriers to confident adoption.

• 2025-Q2: Capability expansion and research advancement coincided with production reliability deterioration. Morphic Studio launched production Video Outpainting extending the practice beyond static images (May 2025). Research progress accelerated with peer-reviewed papers on Vision Transformer efficiencies, color consistency fixes addressing blending artifacts, and human-feedback fine-tuning for medical imaging accuracy. However, production-stage failures accumulated: DALL-E 2 outpainting experienced complete failure (April 2025) affecting commercial users with accumulated credits; Adobe Generative Fill underwent May 2025 regression breaking reference image workflows and disabling text prompting, compounding prior quality issues. The quarter demonstrates persistent bifurcation: academic and domain-specific research advancing sophisticated methodologies and expanding to video, while primary commercial platforms (Photoshop, DALL-E) exhibited deepening reliability regressions. The practice remains constrained by production maturity despite widespread feature availability and theoretical advancement.

• 2025-Q3: Enterprise cloud integration accelerated alongside continued desktop-platform reliability issues and academic maturity consolidation. AWS Bedrock announced Stability AI Image Services GA with production Inpaint and Outpaint tools (September 2025), while Azure OpenAI expanded support with GPT-image-1 offering Precision Inpainting capabilities. A major peer-reviewed survey from Shenzhen Institute, Adobe, and Apple synthesized 100+ research papers and proposed new evaluation frameworks (September 2025). However, Adobe Photoshop Generative Fill suffered critical July 2025 regression blocking portrait edits with overzealous content filtering, while concurrent technical critiques documented persistent weaknesses in OpenAI's inpainting including texture fidelity issues and contextual misunderstanding. The quarter illustrates asymmetric development: enterprise cloud platforms and research sophistication advancing while primary consumer tools remain reliability-constrained.

• 2025-Q4: Enterprise cloud consolidation continued alongside persistent desktop-platform service instability and cybersecurity research advancement. Adobe's Q4 FY25 financial reporting confirmed ecosystem momentum with 30%+ AI-driven ARR and Generative Credits consumption tripling, signaling sustained adoption at scale. However, professional users documented widespread Generative Fill access failures and regional availability blocks (November 2025), indicating infrastructure resilience gaps despite public GA status. Research broadened into synthetic image authentication—arXiv studies evaluated detector robustness against inpainting manipulations, signaling maturity concerns about forensic attribution and deepfake implications. Open-source communities sustained active problem-solving through Stable Diffusion inpainting guides and DALL-E outpainting tutorials (October-December 2025), demonstrating continued developer confidence despite primary platform instability. The quarter exemplifies the practice's defining maturity asymmetry: financial adoption metrics and enterprise cloud deployment advancing strongly, theoretical research sophistication expanding into forensics and security, while primary consumer/professional tools (Photoshop, legacy DALL-E APIs) remain fragile, access-restricted, and requiring persistent workarounds from practitioners—indicating that feature ubiquity and reliability remain fundamentally decoupled.

• 2026-Jan: Feature iteration and quality recovery attempts coincided with persistent reliability challenges and security research advances. Adobe announced upgraded Firefly-powered Generative Fill and Generative Expand (January 27) with 2K resolution, sharper detail, and fewer artifacts, representing vendor response to prior quality concerns. However, deployment barriers remained substantial: technical analysis documented 3-layer gating complexity (identity, infrastructure, intent) explaining grayed-out features, while regional availability blocks persisted through January 2026. Stable Diffusion ecosystem demonstrated continued dominance with 80% market share and 12.59 billion images produced as of 2024. Critical security research (arXiv, January 30) demonstrated state-of-the-art AI-generated image detectors fail dramatically on inpainted content (91%→55% accuracy), exposing detection gaps and deepfake risks. Adoption metrics confirm continued high usage—2/3 Photoshop beta users employ generative AI daily and Generative Fill ranks among five most-used features—yet independent critical assessment documented persistent production limitations including context blindness, texture quality gaps, and hybrid human-AI workflows becoming standard. The month illustrates the practice's core tension: continued feature investment and usage scaling offset by reliability constraints, security vulnerabilities in detection, and production quality gaps forcing reintroduction of human editorial oversight.

• 2026-Feb: Production deployment matured unevenly across platforms alongside critical research and practitioner assessment of limitations. Adobe's 2026 Generative Fill deployment achieved meaningful adoption with field reporting of 40% productivity gains in background replacement workflows on real client work; however, quality limitations persisted with faces landing at 60% realism threshold, indicating incremental rather than transformative improvement. Research advanced technical solutions: CVPR 2024 workshop papers proposed object-aware background generation reducing object extension artifacts by 3.6x versus prior models, addressing known outpainting limitations through improved model architecture. Critical practitioner assessments highlighted production barriers: independent hands-on testing of 30+ tools documented persistent accuracy and face quality limitations requiring human editorial judgment; systems engineering analysis documented architectural trade-offs between blending quality and performance, emphasizing seaming and texture matching remain unsolved. Significant negative signal emerged from professional communities: January 2026 survey of Association of Photographers members reported 58% lost commissioned work to generative AI with average financial losses up 142% to £34,900 per person, indicating both adoption acceleration and economic displacement constraining professional adoption. The month demonstrates the practice at an inflection point: vendor quality iteration and research advancement continuing while professional adoption remains constrained by reliability gaps, quality limitations, and economic disruption in creator communities—the defining pattern of the leading-edge tier where capability availability exceeds production maturity and professional confidence.

• 2026-Mar/Apr: Enterprise cloud consolidation and agentic vendor offerings accelerated alongside forensic research exposing deployment reliability gaps. Adobe released major March 2026 updates with Generative Fill and Expand moving to 2K resolution and support for 25+ third-party models; in April, the Firefly AI Assistant (Project Moonlight) entered public beta with agentic orchestration across Creative Cloud apps, automating multi-step inpainting and outpainting sequences and signalling evolution toward AI-coordinated creative pipelines. Late April brought Photoshop v27.6 with redesigned Generative Fill UI, multi-model picker (Firefly, Gemini, third-party), and stability refinements (crash fixes, panel responsiveness) demonstrating production maturity through iterative polish. AWS Bedrock completed Stability AI Image Services GA on April 28, launching 13 specialized inpainting/outpainting tools (Inpaint, Outpaint, Erase Object, Search and Replace) with Stable Diffusion 3.5, achieving enterprise cloud-native inpainting at scale across AWS infrastructure. Quarterly Photoshop updates (v27.2–27.5) showed active model refinement; Azure OpenAI confirmed production inpainting support across model generations (GPT-Image-2, GPT-Image-1.5, GPT-Image-1, GPT-Image-1-Mini), providing enterprise teams stable documented deployment paths. Technical research matured production optimization: CVPR 2026 papers (InverFill) addressed the speed-quality tradeoff through semantic-aware noise initialization; CVPRW2026 (DiffusionPrint) proposed specialized forensic fingerprinting for diffusion-inpainted regions, confirming inpainting has matured enough to require tailored detection architectures; independent study (April 28) documented gpt-image-2 inpainting at production scale for document editing with sub-second latency. However, critical commercial deployment barriers persisted and new evidence documented them: professional retouchers identified 1024×1024 native resolution cap with forced upscaling destroying fabric texture synthesis; technical analysis documented edge artifacts, color bleeding, and texture mismatches constraining professional production workflows. Production case study (April 24) by Senior Art Director on 10 commercial projects confirmed Photoshop+Firefly production maturity for pixel-perfect product integration with commercial legal indemnification, achieving 3B+ Firefly images generated; however, documented persistent productivity gaps and quality constraints limiting broad professional adoption. Market adoption metrics confirmed mainstream integration: web traffic rankings (February 2026) showed millions monthly users across multiple outpainting platforms; AI image-editor market growing 15.7% CAGR to $8.9B by 2034, with 99% of creatives using generative AI, yet professional economic displacement persisted (58% photographers lost work, average losses up 142% to £34,900/person). The period exemplifies leading-edge maturity at inflection: vendor platform consolidation (Adobe agentic, AWS/Azure enterprise deployment), professional adoption confirmed in specific segments (product photography, commercial campaign workflows), and enterprise cloud delivery advancing—yet quality limitations (resolution, texture synthesis, edge handling), documented professional barriers (required manual correction at retailers 60-80%), and economic displacement in professional photographer communities continuing to constrain tier advancement despite universal feature availability.

• 2026-May: Ecosystem maturation and agentic adoption reinforced leading-edge status without tier advancement signals. Adobe Firefly AI Assistant public beta (April 27) expanded agentic inpainting orchestration across Creative Cloud to mainstream access; a May 2026 Firefly model update delivered sharper resolution, improved photorealism, and stronger prompt understanding, with practitioner-validated workflows documented reducing revision loops. FAL.ai industry report confirmed Black Forest Labs Flux Kontext and Qwen Image Edit as production-quality launches within a 985-endpoint ecosystem, characterising 2025 as a year in which "image editing capabilities revitalized." Consumer-scale outpainting reached 10M+ users via Getimg.ai's commercial platform supporting 9 aspect ratios with batch operations and commercial licensing. Professional post-production assessment documented the reliability ceiling: culling, colour correction, and background isolation automate reliably, but skin textures, deformed hands, jewellery fidelity, and ghost mannequin compositing persistently fail, with hybrid AI-human workflows and mandatory QC gates now the documented production standard. Adoption metrics confirmed momentum: DAIS government-funded report showed 86% of young digital creators adopted image AI for editing/alterations (49% daily, 34% task-specific). However, no evidence of tier-advancing quality breakthroughs emerged: resolution limitations persisted (1024×1024 native), professional economic displacement continued (58% of photographers lost work, average losses £34,900), and commercial-scale retailers (Target, H&M, Amazon) continued requiring manual correction on 60-80% of outputs — confirming ecosystem breadth and adoption depth without breakthrough capability advancement.

• 2026-Jun: Foundation model specialization and agentic research advances marked the period, alongside quantified enterprise ROI at multi-sector scale. Black Forest Labs released Flux Fill GA, a purpose-built inpainting model with context-aware generation and natural texture continuation available via web, HuggingFace, and API—signalling specialized foundation model maturity beyond Stable Diffusion dominance. FLUX.1 Kontext (also GA, fal.ai June 24) added text-guided context-aware editing with a 21M+ run-count ecosystem on Replicate confirming production-scale deployment breadth. Adobe Research presented RetouchIQ at CVPR 2026, an agentic 7B vision-language policy model for natural-language image editing outperforming diffusion baselines (score improved 6.89→7.51), while Adobe filed a patent addressing the known edge-artifact failure mode in user-drawn mask inpainting via two-pass mask relaxation; Adobe's HiLo-Token optimization delivered 1.33-3.13× production speedup with 33% infrastructure cost reduction validated on 110M+ Photoshop requests. Moebius (ECCV 2026) demonstrated 54× parameter reduction (0.22B vs 11.9B) at 15× faster inference matching FLUX.1-Fill-Dev quality, opening a credible edge-deployment path. Multi-sector enterprise ROI was documented at quantified scale: Crystal Method case study showed advertising workflows achieving 30-60% cost reduction with 1-week campaign cycles and interior brands cutting studio shooting 50%; preference-alignment research (Follow-Your-Preference++, UC Santa Barbara) demonstrated improvements across 9 reward models and 2 architectures via direct preference optimization, indicating active methodology advancement. PaintBench's documented 17.1% mIoU ceiling on precise visual editing remained the structural quality constraint, maintaining the practice in leading-edge tier despite broad deployment.

• 2026-Aug: Firefly crossed $300M ARR in Q2 2026 (50% QoQ growth) with its AI Assistant reaching GA (June 18) to orchestrate multi-step inpainting/outpainting workflows across Creative Cloud; 75% of 16,000 surveyed creators describe AI editing as integrated or essential. Enterprise adoption accelerated: 72% of Fortune 500 design teams integrated Firefly into production workflows; Adobe's internal marketing deployment of Firefly/GenStudio delivered 33% faster production and 2x output volume, reinforcing an enterprise-internal ROI signal beyond customer-facing case studies; FLUX.2 reached GA on Azure Foundry with multi-reference editing; Black Forest Labs released commercial FLUX.1 [pro] Fill with explicit pricing ($0.05/megapixel). Yet adoption maturity remains constrained by persistent quality and reliability gaps: 57% of creator outputs still require moderate-to-extensive editing before publication; multi-turn editing exhibits cumulative quality degradation (prompt-based edits drift composition by iteration 4, mask-based edits hold structure but show tonal drift across 8+ sequential steps); 90% of enterprise AI initiatives stall at pilot phase despite universal tool adoption (99% of creatives use generative AI, but only 5.5% achieve meaningful ROI). Market and platform signals confirm leading-edge status without tier advancement: ecosystem consolidation around Firefly, FLUX.2, and commercial APIs; agentic orchestration now standard; consumer authenticity backlash persisting (340% increase in "not real" flags, 22% decline in save rates for AI-smoothed content).