Banana100: Breaking NR-IQA Metrics by 100 Iterative Image Replications with Nano Banana Pro

This is an academic paper (3 days old, 0 stars, 5 forks indicating early arxiv sharing) that identifies a specific vulnerability in how multi-turn image editing accumulates artifacts and defeats no-reference image quality assessment metrics. The contribution is empirical and analytical rather than a new tool or framework. DEFENSIBILITY (2/10): No users, no production artifact, purely a research finding with no buildable system. The code (if released) would be reference implementation only—demonstrating the vulnerability, not solving it. Five forks suggest academic interest but zero velocity indicates no active development or adoption. PLATFORM_DOMINATION_RISK (medium): This vulnerability affects image editing features in Claude, GPT-4V, and similar multimodal systems. Platforms may internalize the findings and implement quality guardrails to mitigate artifact accumulation. However, the finding is narrow—it's a failure mode analysis, not a replacement capability—so absorption risk is moderate rather than high. MARKET_CONSOLIDATION_RISK (low): No incumbent market for 'IQA metric vulnerability discovery.' The finding may inform research directions but doesn't threaten any specific product. Academic papers on metric failure modes are typically cited rather than productized. DISPLACEMENT_HORIZON (3+ years): The vulnerability is niche—relevant to multimodal agent developers and image quality researchers. Platforms may fix it slowly (adding rejection thresholds, iterative quality monitoring). The paper itself won't be displaced because it's a specific empirical finding, but any mitigation it inspires will emerge slowly across the industry. TECH_STACK: Paper-based, likely uses standard CV libraries and existing LLMs for the agent backbone. No novel infrastructure. COMPOSABILITY (theoretical): The finding is conceptual; it's an observation about failure modes, not a reusable component. Any mitigation would require engineering by platform vendors. IMPLEMENTATION_DEPTH (reference_implementation): Likely includes code to reproduce the degradation across iterations, but this is validation code for the claim, not a production tool. NOVELTY (novel_combination): Combines known elements—iterative editing, IQA metrics, LLM agents—but the specific finding (artifact accumulation in multi-turn editing breaks NR-IQA metrics) appears original to this work. The 'Banana100' benchmark (100 iterative edits) is a clever measurement protocol but not a breakthrough methodologically.