Cerca nel blog

2023/11/18

Generare immagini con ChatGPT 4 per illustrare un articolo

Numerosi lettori mi hanno chiesto ragguagli su come ho usato ChatGPT per generare l’immagine di testa di questo articolo sullo stalking tramite AirDrop sui dispositivi Apple. L’immagine è questa:

Immagine generata da DALL-E tramite ChatGPT.

Ho un account a pagamento su OpenAI, e quindi ho accesso alle funzioni di generazione di immagini. Per prima cosa ho chiesto a ChatGPT di generare un’immagine usando questo prompt, inventato e scritto di getto da me:

generate a 16:9 photorealistic image of a woman called Dany on a train. She is looking at her Apple mobile phone with an alarmed expression. The train compartment is lit by a cold, greenish light and the mood is foreboding

Siccome mi ero dimenticato di selezionare la versione 4 di ChatGPT, il servizio mi ha risposto dicendo che non poteva generare immagini ma poteva aiutarmi a visualizzare la scena usando una descrizione dettagliata:

I'm sorry for any confusion, but I can't create images or visual content. However, I can help you visualize the scene through a detailed description. Imagine this:

The scene unfolds in a train compartment, bathed in an eerie, greenish light that casts long, angular shadows across the seats. The light, cold and almost clinical, seems to seep through the windows, tinting the world outside with a ghostly hue.

In the midst of this unsettling ambiance sits Dany, a woman whose features are etched with a mix of concentration and alarm. She's in her late twenties, with a sharp, intelligent face framed by shoulder-length hair that seems to absorb the strange light, giving it an almost ethereal quality.

Dany's eyes, wide and vivid, are locked onto the screen of her Apple mobile phone. The phone, modern and sleek, contrasts starkly with the outdated, slightly run-down interior of the train. Her brows are furrowed, lips parted slightly as if she's on the verge of speaking or gasping, reacting to something startling or unnerving on the screen.

Her attire is a blend of casual and professional – perhaps a smart blouse paired with a comfortable jacket, suggesting she might be returning from work or an important meeting. Despite the simplicity of her clothes, there's an air of sophistication about her, a sense of someone accustomed to navigating complex, high-stakes environments.

The rest of the train compartment is sparsely populated. A few other passengers are scattered around, absorbed in their own worlds, seemingly oblivious to the cold light and the tension emanating from Dany. Their indifference adds to the isolation and foreboding atmosphere of the scene.

The train itself hums and rattles along the tracks, a steady, rhythmic sound that underlies the scene. Outside, the world rushes by in a blur of darkness and occasional lights, further emphasizing the sense of movement and the passage of time.

This setting, with its blend of the mundane and the eerie, frames Dany as a central figure, caught in a moment of revelation or shock, isolated yet surrounded by the anonymous company of her fellow passengers. The cold, greenish light not only sets the mood but also seems to reflect Dany's inner turmoil, making the scene both realistic and charged with an undercurrent of suspense.

A quel punto sono passato alla versione 4 di ChatGPT e gli ho chiesto di usare questa descrizione come prompt, dandogli il comando “generate a photorealistic image using this description, 16:9 format” e ottenendo l’immagine che ho usato nell’articolo:

Immagine generata da DALL-E tramite ChatGPT.

Oggi gli ho chiesto di generare qualche alternativa, visto che alcuni commenti all’articolo originale notavano la mancanza apparente di un corridoio di accesso (c’è, ma è molto spostato sulla destra di chi guarda). Questo è il risultato con la richiesta “redraw this image but with a center aisle in the train car”.

Immagine generata da DALL-E tramite ChatGPT.

Però così sembra un vagone da metropolitana, non da treno, e quindi ho chiesto “redraw but with a long distance train car. Do not use a subway car”. Questo è il risultato:


Immagine generata da DALL-E tramite ChatGPT.

Il livello di dettaglio che si può chiedere a ChatGPT è notevole. Per esempio, ho chiesto di ridisegnare questa scena usando uno stile più fotorealistico e poi di aggiungere sul sedile una borsa che reca il logo “DANY” (ricordando che fino a poco tempo fa i generatori di immagini avevano molte difficoltà a generare lettere o scritte), e ho ottenuto questo:

Immagine generata da DALL-E tramite ChatGPT.

Immagine generata da DALL-E tramite ChatGPT.

Cambiare stile, umore del soggetto e altri dettagli è altrettanto facile:“regenerate the last image with warm sunlight, at sunset, streaming in from the train windows. The woman is smiling and wearing a police uniform. The logo on the bag must say "DANY"” (è importante ricordargli le scritte, altrimenti sbaglia).

Immagine generata da DALL-E tramite ChatGPT.

Direi che questo livello di capacità è piuttosto inquietante. Consiglio di non perdere tempo con ChatGPT 3, la versione gratuita: se volete imparare a usare questi nuovi strumenti, investite in un abbonamento. So che OpenAI ha chiuso gli abbonamenti, ma è una misura temporanea.

Nessun commento: