In the 4o ImgGen best images thread, I noted that I’d love to be able to recreate a DALL-E 3 image’s look in 4o ImgGen by uploading a DALL-E 3 image, getting the AI to provide a prompt that would recreate it in the new model, and then tweaking it with new subject matter.
Aha !!! We have proof of concept !!
Details: I fed ChatGPT 4.5 an image created by DALL-E 3. I asked it to create an image just like it. It got close. I had it adjust the color palette and contrast a bit. It got 90+% of the way there. [ Close enough that I could use Photoshop to get another few % closer. Should get even better with practice.] I then had the AI show me its prompt. I then used that prompt to make a new image in that style. Voila !!
I’d share the ChatGPT convo, it’s fascinating, but conversations with uploaded images can not yet be shared. As soon as that changes, I’ll add a share link to this post.
In the meantime:
Here’s the original script that worked with DALL-E [image shown as Fig. 1]:
“A vintage-style portrait mode illustration with aspect mode 1024x1792. The top half of the image shows a snow-filled stormy sky. The top of the bottom half shows a steep mountain hill, surrounded by forest, with Mt. Shasta in the background. Snow is swirling and blowing. The bottom of the bottom half shows exactly two adorable, laughing baby emperor penguins dressed in colorful winter clothing, sledding down a steep snowy hill amidst a heavy swirling snowstorm, grabbing one another for dear life. Each penguin is adorned in vibrant vintage winter coats, knit caps, woolen scarves, warm pants, thick knit socks, and boots. They are lying prone on classic vintage Flexible Flyer wooden sleds with distinctive red runners and a large, sturdy wooden cross-member at the front for steering. The cross-member is thick, flat, and matches the width and thickness of the sled’s wooden slats. It extends about 6 inches beyond the sled’s rails on each side, designed for steering by shifting it left or right with hands or feet. The cross-member is clearly detailed, emphasizing its solid wood texture and functional design. The hill is encircled by a snow-covered forest of towering pine trees, dusted with fresh snow. In the background, the majestic higher reaches of Mt. Shasta rise, partially veiled by snow clouds. It’s early morning with a golden glow filtering through patches of blue sky. Heavy snow is falling, creating a magical winter atmosphere. The illustration features delicate ink outlines, a muted color palette, and a nostalgic aesthetic reminiscent of early 20th-century detailed illustrations for children. Created using: enhanced hand-drawn textures, fine ink detailing, watercolor, tempera, colored oil pencils, soft shading, nostalgic color grading, vintage children’s illustration aesthetics, detailed fabric textures, early 20th-century print design, warm and gentle lighting, high-definition quality, natural aesthetic, with no framing elements, words, letters, or banners.”
It did NOT work well with 4oImageGen. [Image shown as Fig. 2]
Here’s the derived.by.AI prompt, tweaked only for different aspect ratio and species: [Image shown as Fig. 3] :
“A traditional illustration, widescreen, likely created using colored pencil and pastel techniques, depicting two adorable baby eagles joyfully sledding down a snowy mountain slope. The baby eagles are wearing vibrant winter clothing—one in bright blue and the other in rich red, both accented by colorful scarves and knitted hats with fluffy pompoms. The scene is brightly lit with cool noontime daylight, high contrast, and a fresh, clear atmosphere. Snowy mountain peaks rise majestically in the background, surrounded by a serene landscape filled with snow-covered evergreen trees. The illustration should feature sharp, crisp details, lively expressions, bold and rich colors emphasizing blues, reds, and whites, with excellent contrast and a playful, joyful atmosphere reminiscent of vintage-inspired winter holiday illustrations.”
The smarts of the new model are a wonderful thing.
Also: since I’ve been working on the new model with ChatGPT 4.5, creating lots, it’s about to shut me down until April 2nd. I asked it whether this technique would work with 4o, and it said it should do an even better job, since it knows more about images than 4.5. Off to check that out.
