InkIdeator: AI-Driven Design Ideation

• InkIdeator is a multi-modal AI-powered ideation system that fuses structured cultural annotations with text-to-image synthesis to support creative design, especially in Chinese painting.
• It leverages a richly annotated dataset of 16,315 images and expert-in-loop workflows to ensure accurate genre labeling and semantic clustering for effective ideation.
• The platform integrates interactive tools like symbol panels, moodboards, and image generation features to facilitate rapid ideation and visual draft creation.

InkIdeator is a multi-modal AI-powered ideation support system designed to augment creative exploration and early-stage design in visual domains. The platform integrates structured cultural annotation, knowledge-driven suggestion, and text-to-image synthesis to enable designers—particularly those working with culturally rich styles such as Chinese painting—to efficiently traverse reference collections, synthesize dimensional design intent, and visualize novel compositions within an interactive digital environment. Its core methodology leverages large multi-modal LLMs (MLLMs), expert-in-the-loop annotation workflows, and external generative backends to link conceptual, emotional, compositional, and stylistic information in both guided search and generative synthesis.

1. Data Foundation and Multi-Dimensional Annotation

InkIdeator’s backbone is a curated, extensively annotated dataset of 16,315 Chinese painting images scraped from major social platforms, each coupled with title, author, and description metadata (Wu et al., 26 Jan 2026). The pipeline for transforming this image corpus into a rich, multi-dimensional design space is composed of:

• Manual genre labeling and automated classification: An initial subset (n=1,000; balanced Gongbi/Xieyi) was labeled, and a MambaVision-based feature extractor coupled with a shallow neural classifier achieved 88.0% accuracy in genre assignment, producing genre-balanced subsets of 4,670 Gongbi and 11,645 Xieyi paintings.
• Multimodal annotation via LLM: For each painting, a GPT-4o mini model, prompted as an “art expert,” extracts four dimensions:
  • Cultural symbols: Explicit and metaphorical objects (e.g., “lotus → purity”).
  • Emotions: High-level affective states (e.g., “tranquility,” “majesty”).
  • Compositional patterns: Structural layouts (e.g., “S-curved layout”).
  • Style terms: Brushwork vocabulary, color tones, and overall stylistic markers.
• Quality control and semantic clustering: Expert feedback iteratively tunes LLM prompts and output curation. Textual concepts (1,265 symbols; 4,903 emotions) are clustered using “bert-base-chinese” embeddings and k-means (plus manual refinement).

Key statistics of the annotated design space are summarized in Table 2 of (Wu et al., 26 Jan 2026), with hundreds of unique concept clusters per dimension covering tens of thousands of images, supporting downstream search and ideation.

2. System Architecture and Interactive Workflow

InkIdeator's architecture is structured to combine rapid search, multi-modal suggestion, and generative rendering:

• Front-end: Implemented as a React single-page application providing the main interaction panels:
  1. Symbol Association Panel: Accepts a task theme, presents GPT-4o-suggested cultural symbols and poetic references.
  2. Image Library: Enables faceted search via keyword/annotation filters across the annotated corpus.
  3. Moodboard: Users collect images, with attached dimensional “chips” for further analysis.
  4. Image Generation Panel: Orchestrates chip-driven prompt construction and calls generative backends for sketch visualization.

• Back-end: Flask-based server with:

  • Indexed Elasticsearch over the annotated corpus.
  • Wrappers for GPT-4o mini (annotation, suggestion, prompt composition).
  • MidJourney API integration for visual draft generation.
  • Lightweight PyTorch classifier used for initial genre labeling.

Data flow comprises user theme input, LLM-driven symbol suggestion, Elasticsearch-powered retrieval, semantic chip selection, compositional prompt assembly, and image synthesis (with iterative refinement).

3. Knowledge-Driven Ideation Support

The distinguishing feature of InkIdeator is its symbolic and dimensional ideation support pipeline:

• Cultural symbol suggestion: GPT-4o mini, seeded with a high-coverage few-shot set (20 curated theme→symbol mappings), returns 8–12 symbols per theme with log-probability ranking; empirical user study relevance median is 4.57/5.
• Dimensional keyword presentation: Each image’s annotation chips (symbols, emotions, composition, style) are pre-computed and surfaced for arbitrary user combination.
• Synthesis of dimensionally guided prompts: Selected chips are collapsed into a verbose, design-intention-formatted text prompt, ensuring narrative coherence and contextual domain fidelity.
• Visual draft generation: MidJourney is called with these prompts, optionally including reference imagery for style anchoring. No fine-tuning is performed; the gain in output relevance is attributed solely to prompt enrichment.

User interaction is cyclical and designer-driven: symbol selection guides search, moodboarding and chip selection drive generative input, outputs are further re-chipped for refined cycles, supporting high-frequency re-annotation and exploration.

4. User Study, Evaluation, and Measured Impact

A within-subjects counterbalanced study with N=12 (two 35-minute ideation tasks each) compared InkIdeator versus a baseline (identical image library, but no extracted tags or symbol panel, free ChatGPT/MidJourney access). Principal findings (Wu et al., 26 Jan 2026):

• Idea clarity and appeal: No significant difference (t(11)=0.68, p=0.51).
• Expressiveness, exploration, immersion, and efficiency:
  • Expressiveness: Z=–2.360, p=0.018 (Wilcoxon).
  • Exploration: Z=–2.887, p=0.004.
  • Immersion: Z=–2.228, p=0.026.
  • Organized search/retrieval: Z=–2.724, p=0.006.
  • Extract elements: Z=–2.153, p=0.031.
  • Fast idea-to-visual: Z=–2.310, p=0.021.
• Subjective panel usefulness (1–7 scale):
  • Symbol association: 5.00 (SD 1.28)
  • Image library: 4.92 (1.68)
  • Moodboard: 5.58 (0.90)
  • Image generation: 5.42 (0.90)

Extended case studies with professional painters confirm that tag-driven workflows help concretize narrative details and enable rapid ideation beyond habitual visual patterns.

5. Implications and Broader Applicability

InkIdeator demonstrates the utility of symbolically grounded, multi-dimensional annotation and retrieval in lowering the barrier for novice cultural design exploration while providing experts with high-density compositional inspiration. Its approach amplifies, but can also inadvertently distort, cultural motifs: mis-annotated or LLM-hallucinated tags can propagate errors or introduce unintended stylistic bias, with the risk particularly acute for users lacking domain expertise (Wu et al., 26 Jan 2026).

Recommended co-creation practices include surfacing of validated definitions, clear content attribution, iterative human reflection enforced via UI nudging, and always branding AI-generated images to avoid misuse.

A plausible implication is that such architecture can be adapted to other culturally rich design domains (e.g., ukiyo-e, Western painting) via swap-in annotation pipelines and retrained LLM prompt sets; current limitations include absent sketch-canvas integration and occasional generative hallucination.

6. System Limitations and Future Directions

Noted limitations of InkIdeator in its current implementation include:

• Limited real-world task fidelity: Brief evaluation sessions and small N may not fully capture real-world, large-scale design workflows or downstream creative impact.
• Baseline comparison scope: The adopted “baseline” does not exhaustively represent all real-world design search strategies.
• Absence of in-situ digital sketching: No active integration with a digital canvas for sketch-based refinement; such addition is flagged as important future work.
• LLM-generated poetry: Occasional hallucination in classical poem synthesis for visual augmentations.
• Culturally specific generalization: Robustness across non-Chinese art forms is contingent on annotation schema adaptation.

Scalable extension recommendations encompass integration of sketch-based user input, incorporation of hybrid neural-retrieval poetic references, continuous learning from user correction data, and contextualization for further visual traditions by swapping the cultural symbol/emotion knowledge base.

7. Comparative Context and Potential Extensions

InkIdeator’s explicit focus on culturally structured, multi-dimensional ideation fills a gap not addressed by generic text-to-image pipelines or style transfer tools. In contrast to systems such as Inkspire (sketch-driven analogical ideation with ControlNet and scaffolding) (Lin et al., 30 Jan 2025), or CICADA (vector-based, CLIP-optimized, co-creative sketch agents) (Ibarrola et al., 2022), InkIdeator demonstrates the value of high-granularity, semantic annotation and symbol suggestion in design tasks rooted in cultural and historic tradition. This suggests a productive synthesis between symbolic annotation, analogical suggestion, and multi-modal neural generation models for cross-domain creative support platforms.

---

Key Reference:

• "InkIdeator: Supporting Chinese-Style Visual Design Ideation via AI-Infused Exploration of Chinese Paintings" (Wu et al., 26 Jan 2026)