A few things I learned about using DALL·E

In this post, I run some experiments with DALL·E, including:

  • lighting
  • perspective
  • movie posters
  • variation chaining
16 Likes

Great writeup. I had to do a reverse image search on that Shrimp Logo because it looks so good. The use cases of DALL-E appear remarkable.

I like your note on variation chaining and particularly the first prediction. I’ve seen a post on Twitter where a user detailed how the specific output was reached. e.g. “[Input Prompt] → [Variation] → [Variation] → [Inpainting and Prompt]”, and I could see something like that becoming standard DALL-E notation.

1 Like

Lucky you! :wink:

Nice write-up. Thanks for sharing.

1 Like

Indeed, the creative flexibility granted to creators using variations and prompting off of root images is quite a big deal

1 Like

cool the album cover :slight_smile: :white_check_mark: @bram

1 Like

I’ve collected some exotic images, could you try Variations on them and post snapshots of the 10 DALL-E 2 makes?

And could you try:

A glossy black castle’s front gate with gold swords across it. On both sides of the gate is red roses in glossy black gardens, and gold Egyptian pots with handles. Smoke is coming out through the gate. Glossy black stairs lead up to the gate. A few large ghostly glowing light blue Japanese letters float in front of the gate. A fountain is releasing sparkling water onto the floor past the gate. Ultra 4K realistic photograph.

Macro view of a video game case titled “Metal World” sitting against a white wall on a white table near a window. Rated E. The game case is grey rimmed. On the cover of the game case shows only a shallow height 3D metal maze world with lots of dead ends which is where black pipes are. There is metallic creatures throughout the metal maze. Throughout the maze is horizontal red laser beams that act as gates. Some dead ends of the metal maze have small metal castles. Ultra 4K realistic photograph.

A glossy black mega world maze underground surrounded by lava. White standing blackbucks guard the entrance gate. Large gold keys float throughout the maze. Around the maze is glossy black castles with many levels that wrap around. Ultra 4K realistic photography.

A white robot with gold hands with finger tips made of nanobots and a tall head stands in the middle of a glossy white room on a red carpet looking at the camera with blue shining eyes. Thick black electrical wires are connected to its head, arms, and legs, and are releasing smoke. A tall gold rimmed scanner is looped around the robot scanning it. Several floating cameras surround the robot watching it. Nanobots and a gravitational field can be seen around the robot. Ultra 4K realistic photograph.

Is root image the same an init image in diffusion ?

makes such a hell of a change to see a really well thought through piece like this :slight_smile:

I’m STILL waiting for access :confused:

The use of inpainting to expand images is incredible.

This expansion of The Scream is one of my favorite outputs yet.

It continues to be astounding what users are able to come up with. The use-cases for DALL-E seem to keep expanding.

2 Likes

These are so incredible , I literally can’t wait to try to flipping perspectives and lighting too

1 Like

I’m really intrigued and excited about this software. As a graphic artist who struggles with ADHD, this would help me execute my thought process.

2 Likes

Thank you for posting this, I was able to apply the learnings in your post to a different text-to-image model which significantly improved the results.

For context, the model I mentioned above is really bad with human faces and animals. But I am happy with the results I am getting generating artworks such as paintings. It does requires a rather long and detailed prompt for which I have been leveraging the completions api and text-davinci model. I can only imagine how much better it would work with DALL-E 2.

1 Like

Really keen to get access to DALL-E2. Is there any info on the price per image generated? Thanks

I’m curious about the continuity in the animation chain. Could you try generating a sprite sheet? 1024px size isn’t bad at all for producing keyframes. Especially pixelart style. A prompt like this: “A pixel-art style sprite sheet featuring a flaming sword swing animation of a half-penguin, half-orc character in a cloak.” (I’ve no idea why :penguin: )

Maybe even a picture that is filled in the first frame and left the other frames blank can be used to use the character you designed yourself. :thinking:

Animation in this case may not be the traditional “character moves across screen” but instead an “evolution” of pictures.

For example: https://twitter.com/model_mechanic/status/1518515033025642496?s=20&t=di8AMigM1jA5SCh9eHP1BA

2 Likes