SDXL SiFi Movie Knowledge

SDXL SiFi Movie Knowledge

Today I found myself a little bit bored for about a second and typed 'Star Wars' into Stable Diffusion XL just to see what would happen. The model I was using seemed to have some reference training on the movie Star Wars as the image produced was clearly trained on the franchise. So, I decided to give a few movies a try. I know this is a truly pointless exercise, but sometimes that is okay. Here are some of the results.

Stable Diffusion XL

Setup and settings

For this setup I am using my primary development server (R9 7590X 128GB DDR5, RTX3060 12GB). I am using the model CNCXL v 1.0. To interface with Stable Diffusion I am running Automatic1111's WebUI in a docker container.

The prompt is formatted as "Still movie frame of {{Name of Movie}}, 35mm film, dark theme". The sampler/solver is DPM 3++ SDE Exponential. Resolution is 1024x576. CFG Scale is 3- 7. No negative prompt or refiner. And I am running 2 batches of 2 images each time.

Star Wars

According to Stable Diffusion

Honestly, I am pretty impressed. For current open source AI the logo is very good. The characters are well represented, and the theme is spot on.

2001: A Space Odyssey

According to Stable Diffusion

Again, the model certainly seems to understand what I am referencing. It follows the theme very well. I can hear the quietness of the film looking at these images.

Blade Runner

According to Stable Diffusion

I'm not as familiar with this movie. I think it got the theme right. Not so sure about the helicopter or other subjects.

Alien

According to Stable Diffusion

As expected the prompt 'Alien' is too generic, even with the word movie contained in the prompt. Interestingly, there is some theming of these aliens that is clearly being pulled from the franchise.

Ridley Scott's Alien

According to Stable Diffusion

This prompt seemed to help a bit. But still would need some work. Not very impressed.

The Matrix

According to Stable Diffusion

Seems to have been trained on this franchise as well. As expected it isn't very specific with the glyphs. We also see a bit of merging between the 35mm film and the subject in the last one.

1927's Metropolis by Fritz Lang

According to Stable Diffusion

I was more specific with this prompt. The results are certainly not horrible. Not exactly true to form, but got the theme spot on.

1968's Planet of the Apes

According to Stable Diffusion

I mean, this isn't the worst thing ever, but I'm not very impressed. I have a feeling the franchise is in the training, but pulling out the correct vector seems to need some prompt engineering.

John Carpenter's The Thing from 1982

According to Stable Diffusion

It obviously has been trained on this movie. Nothing is quite accurate but kind of close. And once again it got the logo correct.

1956's Forbidden Planet

According to Stable Diffusion

I don't know how accurate this is. But I do love these images. The retro vibe is amazing.

Spielberg's 1982 film E.T. the Extra-Terrestrial

According to Stable Diffusion

Clearly, ET is not quite right here. But he does look amazing, just off. I don't know why I like these so much. But I have to be honest that they are just not accurate at all.

ET: The house elf?

2004's Eternal Sunshine of the Spotless Mind

According to Stable Diffusion

This is a case of right movie wrong actors. Kind of interesting honestly. I don't have to much to say here as it just isn't very interesting.

Star Trek II: The Wrath of Khan

According to Stable Diffusion

This one kind of just is what it is for me. It obviously knows what I am talking about. However, the images just are not very good. I am not quite sure who's face that is, and the ships are just messed up.

1956's Invasion of the Body Snatchers

According to Stable Diffusion

This one seems a bit confused, 1950s? 1970s? And I don't know what is wrong with the creature but it is just... well...