DALL-E 2 Unbundling
As the DALL-E community grows bigger and a part of our daily lives, multimodal artists and their audiences may seek greater novelty.
Gone are the days where you could just share an image “in the style of van gogh” and get recognition from others for your DALL-E 2 creation.
However, since my first month or so, I’ve been using a process I call, “unbundling” to bring van gogh-like familiarity but with some variation. The benefit of this approach is you get results which are slightly different and appear distinct, maybe even a little bit new, but also subconsciously familiar to your audience. It also allows you alternatively explore the model’s implicit representations we have of existing great works from new angles.
When it comes to the world of Multimodal AI models like DALL-E 2, I would define Unbundling as:
The process of exploding a label of an artistic style, known work, art movement, or universally recognized symbol into a longer description of its various characteristics and components.
The idea is that you take a prompt label such as, “in the style of Van Gogh” and try to describe his style in a few words or sentences. Afterwards, instead of using the broader label, you feed in the unbundled version into DALL-E 2 and get back different but subconsciously familiar results.
To generate unbundled descriptions of DALL-E 2 labels, you could do this by hand, and it may be a great training exercise for DALL-E prompt design too, but I personally just use GPT-3 for this process and save a lot of time this way. I’m going to show an example below.
Unbundling Case Study - Van Gogh’s Starry Night
Without Unbundling
“Illustration of a soccer pitch in the UK at night time with stadium lights on, in the style of starry night by van gogh, digital art”
With Unbundling
“Illustration of a soccer pitch in the UK at night time with stadium lights on. The painting is composed of thick brushstrokes of paint that are applied in a swirl-like pattern. The colors are mostly blues and greens, with some yellow and red. The stars in the night sky are rendered with white dots. The painting has a dreamlike quality, and the overall effect is one of intense emotion. Digital art”
Although you may be limited to GPT-3’s existing knowledge of an art work or movement here, I still think the process produces some interesting results!
Correction Update - July 4, 2022
Looks like I mixed up Picasso with Van Gogh above in the prompts lol regardless, looks like GPT-3 and DALL-E still handled this typo well and ignored and knew what I meant. The systems generally do OK and handle typos well. The results here still look good and I encourage you to try this method yourself (with correct typo-free prompts, of course). I’ve left the typo, incorrect prompts in this article here to preserve the article in its raw format, please keep it in mind though.
I also got a really interesting question on Twitter. This technique isn’t limited to Starry Night by Van Gogh! The main idea is that you can find unbundling prompt inspiration everywhere.
By the way, Jeff has also pointed out that you can ask GPT-3 for emotional unbundling too:
The full prompt he entered into Midjourney was, "Oil Painting of Stonehenge with intense feeling and great poignancy. The characteristics of the painting are its strong colors, its deep shadows, and its overall feeling of great movement. The style of the painting is expressive and deeply emotional."
Correction Update - July 5, 2022
I have updated the post above with accurate prompts and images, for historical reference purposes, the old ones with the typos are still available here below:
Without Unbundling
“Illustration of a soccer pitch in the UK at night time with stadium lights on, in the style of starry night by picasso, digital art”
With Unbundling
“Illustration of a soccer pitch in the UK at night time with stadium lights on. The painting is highly stylized and an abstracted painting that depicts a night scene with a large central sun and smaller moons in a swirling sky. The painting is done in a limited palette of blue and white, with thick brushstrokes that give the impression of movement. Digital art”
Thanks for explaining this. Have been learning how to use GPT3. Now have been transferred to paid user and they gave us some free credits for being Beta Testers. Tried some old prompts. Something has changed, and not for the better. Will have to use GPT3 to tweak the original prompts and see if that helps.