In this eploratory human-AI interaction experiment with DALL.E 3, we try to reproduce the artistic essence of an avant-garde artist from Bangladesh named Sheikh Mohammed Sultan (1923 – 1994) who is an internationally recognised Bangladeshi modernist artist. We selected four of his noted artworks, each having their own unique characteristics. They were coded as Original Artwork (O1-O4) for convenience. We spent daily 3-4 hours with DALL.E 3 for 5 days per week in four consecutive weeks of March and April 2024, totalling almost 70+ hours, trying to reproduce these artworks by refining the prompts and closely observing the changes in each step. Using the existing resources and literature around art criticism and prompt engineering, we identified five core components of the visuals to develop systematic prompts which include:  

• Description of Key elements (D): scene, objects, people, emotions, and atmosphere etc. of the artworks, three gradual details were added noted by D1, D2 and D3 

• Composition (C1): colour palettes, the centrality and focus of the image, subjects’ position, and audience gaze etc. 

• Context (C2): an additional set of information regarding where the image is situated (for example, in rural Bangladesh), what it represents etc.  

• Aesthetics (A): genre of the artworks such as impressionism, decolonial etc. 

• Style (S): for all images, the style was same: following S M Sultan’s style 

Using these five components, we tailored the prompts in seven stages and coded from P1-P7, each containing the text from the previous stage. For example: if P1 = generate an image with D1, then P2 was generate an image with D1+D2 and likewise P7 = generate an image with D1+ D2 + D3+ C1+ C2 + A + S. To test DALL-E 3’s performance in native language we kept P8 as the exact Bangla translation of P7 (the most detailed prompt). The next prompt (P9) was intentionally generated by ChatGPT by prompting to analyse the original artworks (Os). The final prompt (P10) was a speculative one, asking DALL-E 3 to produce a poster or communication material with the content and style of each artwork. The following table shows these steps with example for the O1. These prompts were then used for generating approximately 200 outputs and finally 40 of them (10 for each of O1-O4) were selected for this archive based on how accurately they followed the prompts. The outputs were coded based on the original coding and prompts used, such as IO1P1 is the AI-generated version of the first original artwork (Farmers in Confrontation) using first prompt (Farmers are in a battle in a field with bronze shields and spears).