Documento originale
GPT Image 2 · en
Scene:
A square photoreal image of a museum-scale glass atrium installation that looks like a neural network explaining itself. Use soft daylight, polished concrete, suspended translucent panels, and clean reflections. The camera is slightly elevated, looking into a layered architectural space where the diagram is physical, luminous, and readable without text.
Subject:
A Transformer self-attention process represented as a suspended kinetic sculpture. Rows of small glowing token prisms enter from the left, split into query, key, and value-like light paths, pass through multiple attention heads as nested rings, then recombine into a refined output ribbon on the right. Use color-coded flows, curved cables, transparent matrices, and small abstract annotation plaques made only of symbols, ticks, and dots. The structure should feel academic and meta: an AI model drawing its own internal attention map as a physical exhibit.
Important details:
The composition must make process logic visible: input sequence, parallel attention heads, weighted connections, merging, normalization-like layers, and output. Keep node spacing regular, avoid tangled spaghetti, and maintain clear depth separation between foreground paths and background panels. The image should balance photoreal materials with exact technical-diagram hierarchy. No literal equations are needed; the flow itself should explain the mechanism.
Use case:
Neural Network Self-Explanation single study 1/8, testing a Transformer self-attention process as a photoreal academic installation.
Constraints:
No readable text, no watermark, no logos, preserve exact 1:1 aspect ratio, no real model names, no code snippets, no private chain-of-thought, no random neon chaos, no impossible cable intersections, no generic brain icon, no unreadable clutter.Come usare
- Copia tutto il prompt originale, inclusi i parametri.
- Adatta soggetto, testo e composizione alle tue esigenze.
- Usa uno strumento esterno compatibile; verifica accesso e risultato.
Prima di iniziare
- Controlla se serve un’immagine di riferimento. La compatibilità non prova il modello usato per l’esempio.
- Verifica nomi, numeri e testi; i risultati cambiano con le versioni del modello.
