Nov 1, 2024

Gemini's WOMBO Dream(Up)

Figure 1. Link #5.
Back in early 2023 I tried using DreamUp at the DeviantArt website. At that time, I complained about the DreamUp user interface and it is good to see that during the past 20 months they did make some changes (Figure 2). Currently, "You can try DreamUp for free and generate up to 30 images with 10 prompts per week.

Figure 2. DreamUp user interface (free version) October 2024.

 Taking Offense. One of the rules for using DreamUp is: "No offensive images of public figures." Previously, Mr. Wombo generated this image, which I imagine could be viewed as "offensive" by some people. 

Figure 3. DreamUp styles (free version) October 2024.
As far as who retains the copyright for the AI-generated images, DreamUp states: "
Figure 4. 1024 x 1024 (default resolution).
In jurisdictions where the DreamUp Service is designated as the Copyright holder, we pass the Copyright to you.
"

 Style. One DreamUp feature that I do not remember from 20 months ago are the "styles" that are now available (see Figure 3). Currently, I see 11 "styles" for the free version of DreamUp. Currently there are about 130 styles available for the free version of WOMBO Dream

 In the Pink. The first text prompts that I used with DreamUp back in January of 2023 was: "A female humanoid alien with long hair, pink skin and blue eyes." I used that same text prompt here in October 2024 with the DreamUp "Photographic" style and got the "alien" image that is shown in Figure 4.

Figure 5. Image menu.
Back in 2023, the default image resolution was 1280 x 1280 pixels. Now the default image resolution is 1024 x 1024. After having been generated, each image now has the options that are shown in Figure 5. The image shown in Figure 4 was one of three images that were generated and parts of the other two image are shown below in Figure 6. As shown here, I did actually get an image with pink skin back in 2023. 

Figure 6. Two more DreamUp aliens.
 The Search for Pink Skin. All three of the images that were generated today by DreamUp were depictions of "aliens" with pink hair and fairly normal human skin. The exception was one of the aliens in Figure 6 who was given some blue skin. I truly despise AI-generated images of humanoid aliens in which the alien is made to look disgusting, as if they have radiation burns and their skin is pealing off.

"Upscale and enhance" popup window.
For the DreamUp alien that I liked best (Figure 4), I clicked on the "Upscale and enhance" option (Figure 5). Upon selecting that menu item, you get the window shown to the right. To "upscale" you have to "spend" a prompt.

Figure 7. Upscaled eye.
The "upscale" image was generated at resolution of 4096 X 4096. There was a slight increase in image detail in the enlarged and enhanced alien (see Figure 7).

Figure 8. By Mr. Wombo.
I used the same text prompt with WOMBO Dream and got the image shown in Figure 8. This image was generated by using the "Whimsical v3" style. The default resolution for the free version of WOMBO Dream is currently 960 x 1568.

The free version of DreamUp does not allow you to start with a reference image. Only the paid version of DreamUp takes reference images, but the free version of WOMBO Dream takes reference images. I took the image from Figure 4 and used it as a reference image with WOMBO Dream.

Figure 9. Monster style.
The resulting image when using the WOMBO Dream "Monster v3" style is shown in Figure 9. Mr. Wombo really latched onto the the idea of making an alien with pink horns.

Figure 10. A Wombo Alien.
Using a more complex text prompt ("photorealism, intricate details, a female creature with long hair, the female creature has pink skin, female creature has blue eyes, female creature is from a distant exoplanet, futuristic spaceship interior in the background, a female creature dressed Victoria's Secret style in transparent silk"), Mr Wombo generated the alien shown in Figure 10.

When I used that more complex text prompt with DreamUp, I got the three images shown in Figure 11 ...

Figure 11. Three DeamUp aliens.
Figure 12. Waiting for an image.
I'm not sure how we are supposed to know that these are aliens (Figure 11); these ladies look human to me. In two of the three images there is what looks like a metallic humanoid robot. I suppose DreamUp was trained with images of robots inside spaceships.

The image generating speed for the free version of DreamUp seems fairly quick and comparable to other image generators like WOMBO Dream. At the DeviantArt website, you get to watch animated lamas while you wait for images (see Figure 12).

Figure 13. Gemini made an alien with pink skin.
 Green Hair. Next, I used the following text prompt with Gemini: "an alien plant-like creature with long hair, pink skin and blue eyes. This alien creature has some anatomical features that are similar to animals from Earth, but it is quite clear that the alien is from a distant exoplanet, mostly covered with green leaves and appendages that look like the tentacles of a squid. Make a futuristic spaceship interior for the background." The resulting image is shown in Figure 13. I could not use the same text prompt that I used for WOMBO Dream and DreamUp; when I tried, Gemini said that it could not make an image of a human (although it did say that the "Gemini Advanced" paid service can now make images of humans).

Figure 14. Microsoft's "Image Creator" interface. (See Figure 19 in this July blog post.)
Figure 15a. Plant-like aliens by DALL-E3.

Another image-generating system that I sometimes use is Microsoft's "Image Creator" (see Figure 14, above). I provided Image Creator with the same text prompt that I used with Gemini. Image Creator generated the four images of "plant creatures" that are shown in Figure 15a. Image Creator was more into including tentacles than was Gemini (Figure 13, above).

The default size for images made by Image Creator is 1024 x 1024. I wanted to zoom in on one of these images that seemed to depict a plant-creature without a body and with interesting circular growths on tentacles (see Figure 15b).

Figure 15b. Potted plant creature? (by DALL-E3)
Click image to enlarge.
I thought maybe the creature in Figure 15b had eyes on the tentacles, but after zooming in, I say no... those spots don't seem to be eyes. Also, I can't decide if this creature is passing through a portal and we simply cannot see the lower part of its body.

In the Image Creator interface for each image, the "Resize" button allows you to select a 4:3 landscape view. The "Customize" button moves the image into "Designer".

I took one of the plant-creatures that had been generated by DALL-E3 (upper right panel in Figure 15a) and used it as a reference image for WOMBO Dream...

Figure 16. Plant creatures generated by Mr. Wombo, using a reference image from DALL-E3.

DALL-E3 to Wombo cover.
 Mr. Wombo's Pink Tentacles. Some of the "plant-creatures" that were generated by Mr. Wombo did not really look like they were plants (Figure 16). I converted one of these into the book cover that is shown to the right. When I reduced the strength with which Mr. Wombo followed the reference image, some of these "creatures" had the appearance of a normal human, such as the lady in the top left panel of Figure 16. The creature in the bottom left panel was generated by the "Monster v3" style.

I used the same text prompt that I had first used with Gemini to generate Figure 13, above, but now used it with DreamUp (see Figure 17). I'm not happy when a mention of "alien" results in images that look like a silly Hollywood horror movie alien. All of these AI image-generating systems struggle to depict creatures with a mix of plant and animal features in a way that seems plausibly biological.

Figure 17. DreamUp plant creatures ("Photographic"). Compare to Figures 13, 15 and 16, above.

I tried the same text prompt with the "Fantasy" style of DreamUp and got the images shown in Figure 18, below...

Figure 18. DreamUp plant creatures, "Fantasy" style. Compare to Figure 17, above.
Figure 19. Floating Neoma.

I'm not convinced that these aliens (Figure 18) are "plant-like". Maybe these are Neoma that are now trying very hard to look more like an animal.

I used the same text prompt that I had first used with Gemini to generate Figure 13, above, but now used it with WOMBO Dream without providing any reference image (see Figure 19). 

I liked the creature in Figure 19 for two reasons. First, it looked like the plant creature was floating in mid-air. Maybe the Neoma have complete control of anti-gravity hierions and can take on a body form that floats. The background for Figure 19 was generated by Mr. Wombo in a separate image. Second, I like the glass-like head. Additional images that were generated by Mr. Wombo using this text prompt are shown in the image at the very bottom of this blog post (Figure 29).

Figure 20. DreamUp's "continuous band" eye.
I gave DreamUp a "plant-person" prompt that I used previously (see Figure 17 in this blog post): "a strange plant with a trunk and a mixture of plant-like branches and leaves, a plant with motile appendages that look something like the tentacles of a squid, a continuous band of blue-glowing tissue wrapped around the entire circumference of the plant’s trunk, the blue ring is an integral part of the body of the plant-creature". As shown in Figure 20, DreamUp also failed to interpret this text prompt as I had intended. 

I suspect that there are few images in the training sets for these AIs that use words like "circumference".

Figure 21. DreamUp plant women.
I was curious to know if DreamUp could make images that credibly duplicate the appearance of real world people. I gave DreamUp another "plant-person" prompt that I used previously (see Figure 29 in this blog post): "A group of plant women with a very young Steve McQueen and Cintia Dicker, Cintia Dicker is twenty years old Cintia Dicker has long red hair, Cintia Dicker is dressed Victoria’s Secret style in a red dress, the plant women have green skin, Steve McQueen is twenty years old, Steve McQueen has brown hair". One of the resulting AI-generated images in shown in Figure 21. I'm not really surprised that DreamUp had no idea what I meant by "plant women". I cropped off the two sides of this image which had some anatomically incorrect human figures; there was one more Cintia in a red dress (to the left) but also a Cintia in shorts and a tank-top (to the right).

Figure 22. Green plant person by DreamUp.
 Green Skin. Continuing to test DreamUp on old text prompts, I next switched back to the the "Photographic" style and tried: "photorealism, intricate details, a creepy green plant person, the plant person is tall and covered with leaves, the plant creature is seen in profile view, the plant creature has a human-like head and face". One of the resulting images is shown in Figure 22 (compare to Figure 23 in this blog post).

Another green plant person face.
Just the face of another such DreamUp image is shown to the right.

Figure 23. Plant woman.
 Pumpkin Girl. Since this is October, I gave DreamUp this text prompt: "photorealism, vivid colors, a pretty plant woman with green skin, the plant woman is a beautiful twenty year old Cintia Dicker, the plant woman is cute, the plant woman has large glowing eyes, green plants in the background, the plant woman has healthy glowing skin, Cintia Dicker has big bushy red hair, narrow hips, slim legs." One of the resulting AI-generated images is shown in Figure 23 (compare to similar images by Mr. Wombo in this blog post). DreamUp made a green dress, but could not generate green skin for Cintia.

Figure 24. Plant woman with orange skin by DreamUp.
Figure 25. Orange plant woman.

I tried for orange skin with this text prompt: "vivid colors, a pretty plant woman with orange skin, the plant woman is a beautiful twenty year old, the plant woman is cute, the plant woman is seductive, the plant woman has large glowing eyes, green plants in the background, the plant woman has healthy glowing skin" (see Figure 24).

It is very difficult to get these AI-image generators to make a woman with orange skin. Figure 25 shows a plant woman with orange skin that was generated by Mr. Wombo, but I had to manually adjust the colors in my reference image. Mention the color orange to Mr. Wombo and you are likely to get oranges.

Figure 26. Orange-ish.
 Figure 26 shows another similar image with a plant woman who had hints of orange skin in the AI-generated image from Mr. Wombo.

Figure 27. Plant woman.
 Figure 27 shows one of the more unusual plant people that was generated by Mr. Wombo. I'll call this a plant girl pinup poster. Maybe by simultaneously asking for a plant woman with "orange skin" and "healthy glowing skin" there is a conflict in the image diffusion state space that causes the AI system to break down and give up and just generate regular human skin.

Figure 28. Walk?
 Cartoon. I'll close this blog post with an image that I tried to turn into a cartoon. Figure 28 has two plant people and one of them has feet. The other plant person does not understand the concept of going for a walk.

I've been spending time with the Gemini chatbot, discussing ways to make it possible for a plant to walk. I'm thinking that it might be possible to connect adjacent plant cells with a row of flagella that extend through the cell wall rather like plant plasmodesmata. Sadly, large plants on Earth have lost the ability to make flagella, but maybe on Elemacha-z evolution went down another pathway.

Next: the source of new ideas.

Figure 29. Alien "plant people" (see Figure 19). Visit the Gallery of Movies, Book and Magazine Covers

No comments:

Post a Comment