The color collocation of the design scheme should meet multiple constraints such as functional semantics, visual harmony, accessibility and style consistency at the same time. The existing rule template, image extraction and condition generation methods have structural contradictions between constraint controllability and combination diversity, and the color relationship is always outside the generation process, only for post-output screening. In this paper, the problem is formulated as a sequential decision problem of discrete color combinations under multiple constraints, and the relation-aware diffusion generation framework ColorRel-Diff is proposed. The discrete color token sequence is used as the generation unit, and the mask discrete diffusion is used to make each denoising step explicitly masked discrete diffusion. The color relationship coding module is constructed, and the harmony and contrast relationships are encoded as the edge features of the relationship graph and aggregated into the relationship perception embedding through the graph attention. A constraining-guided denoising mechanism is designed to guide the sampling trajectory by the gradient of relational consistency score, and the feasibility mask is used to impose inviolable token-level restrictions on the hard constraints, and the grading softening strategy is used to deal with constraint conflicts. A three-layer inspection system of distribution level, constraint level and design availability level is constructed, and experiments and ablation are carried out on the basis of increasing constraint strictness. The results show that the proposed method occupies a Pareto-dominant position in the diversity-harmony plane, and the satisfaction rate of accessibility hard constraints remains above 95% under pressure, which is significantly better than the baseline without explicit modeling constraints. The ablation confirms that relational coding and constraint guidance dominate the gains of harmony and constraint compliance respectively.
Selecting a colour palette that suits its design context (application domain, audience, and accessibility profile) affects usability, engagement, and accessibility, yet remains a difficult computational problem. Vision–language models (VLMs) have improved performance on this task but share three limitations: deterministic outputs cannot represent the variance inherent in a design brief, single-scalar training fixes the harmony–accessibility trade-off at training time, and post-hoc context filtering leaves the generator itself context-blind. We argue that these limitations stem from the output representation rather than the encoder. CAIM (Context-Aware Intelligent palette Modelling) represents palettes as a context-conditional Gaussian mixture in CIELAB whose parameters are predicted from frozen vision–language embeddings by a learned head. It refines sampled candidates under five objectives designed to target distinct and complementary mathematical features with reference-direction-based many-objective optimisation and ranks them with a context-conditioned selection rule. On four public datasets and a 3200-scenario context-rich dataset, CAIM achieves up to 23% higher predicted ratings than the strongest baseline, a 22-percentage-point gain in APCA readability over the strongest baseline, and a Pareto hypervolume 34% larger than the convex hull of baseline solutions, at 1.6 s per scenario. Modelling VLM-conditioned generation as a context-conditional distribution is promising for design tasks with inherently uncertain outputs.
Abstract Discreteness of representations is a core commitment of classical generative phonology. However, one of the biggest arguments levelled against discreteness has come from the phenomenon of incomplete neutralisation – where the output of some phonological process on a feature [F] is phonetically distinct from underived [±F]. A sizeable body of work has contended that incomplete neutralisation crucially undermines the sufficiency of discrete representations, arguing instead for the inclusion of gradient representations. In this article, we investigate backness harmony in Uyghur, evaluating recent claims of incomplete neutralisation. Results from two production studies replicate findings from previous work on the language, but we show that, with proper experimental controls, the pattern is derivable without recourse to gradient phonological representations. We then discuss the conceptual and theoretical challenges for gradient representations along with criteria for true phonologically relevant gradience.
چکیده: گسستگی بازنمایی‌ها یکی از تعهدات اصلی واج‌شناسی زایشی کلاسیک است. با این حال، یکی از مهم‌ترین انتقادها به گسستگی، از پدیده خنثی‌سازی ناقص ناشی می‌شود؛ جایی که خروجی برخی فرایندهای واجی بر روی یک ویژگی [F] از لحاظ آوایی با [±F] غیرمشتق تفاوت دارد. حجم قابل توجهی از پژوهش‌ها استدلال کرده‌اند که خنثی‌سازی ناقص، کفایت بازنمایی‌های گسسته را به طور اساسی زیر سؤال می‌برد و در عوض، بر ضرورت گنجاندن بازنمایی‌های گرادیانی تأکید می‌کنند. در این مقاله، هماهنگی پسین در زبان اویغوری را مورد بررسی قرار می‌دهیم و ادعاهای اخیر درباره خنثی‌سازی ناقص را ارزیابی می‌کنیم. نتایج حاصل از دو مطالعه تولیدی، یافته‌های پژوهش‌های پیشین در این زبان را تکرار می‌کند، اما نشان می‌دهیم که با کنترل‌های آزمایشی مناسب، این الگو بدون نیاز به بازنمایی‌های واجی گرادیانی قابل استخراج است. سپس، چالش‌های مفهومی و نظری بازنمایی‌های گرادیانی و همچنین معیارهای گرادیان واقعی مرتبط با واج‌شناسی را مورد بحث قرار می‌دهیم.
