Experiments in AI generated Kerbal art.

[Aelfhe1m]

Some images I created using Automatic 1111 (A1111, a Stable Diffusion based stand-alone art generator).

Like all AI art, these were a bit hit and miss, with even simple prompts sometimes generating something not even remotely related to what I was asking for. But enough turned out halfway decent that I am willing to share them (note these images were previously posted to the "What I did in KSP1 today" thread)

 

[Fizzlebop Smith]

i was having a very difficult time in getting it to produce images without helmets. 


@Aelfhe1m
Do you mind if i use these to train my model?

[Aelfhe1m]

6 minutes ago, Fizzlebop Smith said:

i was having a very difficult time in getting it to produce images without helmets. 


@Aelfhe1m
Do you mind if i use these to train my model?

I don't mind. Not sure what the copyright would be, since I used a premade LORA to generate them (https://civitai.com/models/181378/kerbal-space-program-kerbals)

It does have a very strong bias towards helmets and spacesuits, but by adding them to the negative prompt, I was able to get most images in these batches to generate without them.

An example prompts:

(first image):

Spoiler

1guy, kerbal, Scottish, 1700s dress, outdoors, jungle, coastline, standing on the jungle coastline of south America, 18th century, 1800s,<lora:kerbal:1>
Negative prompt: (helmet), (space suit)
(bad quality, worst quality:1.4), Asian, plain background, steam, breath, motion lines, shuddering, greyscale, (CyberRealistic_Negative:.6)
Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3249766356, Size: 600x900, VAE hash: c6a580b13a, VAE: vae-ft-mse-840000-ema-pruned.safetensors, Clip skip: 2, Lora hashes: "kerbal: 1e54e909b799", TI hashes: "TH: 58af0a72a597", Version: v1.6.0-2-gb1d24c7d

(fifth image - humanish Kerbal girl in red skirt)

Spoiler

vintage, 1girl, kerbal, Scottish, 1700s dress, outdoors, jungle, coastline, standing on the jungle coastline of south America, 18th century, 1800s,<lora:kerbal:1>
Negative prompt: (helmet), (space suit)
(bad quality, worst quality:1.4), Asian, plain background, steam, breath, motion lines, shuddering, greyscale, (CyberRealistic_Negative:.6)
Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3857605576, Size: 600x900, VAE hash: c6a580b13a, VAE: vae-ft-mse-840000-ema-pruned.safetensors, Clip skip: 2, Lora hashes: "kerbal: 1e54e909b799", TI hashes: "TH: 58af0a72a597", Version: v1.6.0-2-gb1d24c7d

(the last image):

Spoiler

1guy, kerbal, Scottish, 1700s dress, outdoors, jungle, coastline, standing on the jungle coastline of south America, 18th century, 1800s,<lora:kerbal:1>
Negative prompt: (helmet), (space suit)
(bad quality, worst quality:1.4), Asian, plain background, steam, breath, motion lines, shuddering, greyscale, (CyberRealistic_Negative:.6)
Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1343213339, Size: 512x768, VAE hash: c6a580b13a, VAE: vae-ft-mse-840000-ema-pruned.safetensors, Clip skip: 2, Lora hashes: "kerbal: 1e54e909b799", TI hashes: "TH: 58af0a72a597", Version: v1.6.0-2-gb1d24c7d

Some of my "odd" results:

Spoiler

 

[Fizzlebop Smith]

The copyright fine as long as you acknowedge the  models. As long as you arent actively trying to reproduce works of a particular artists style.
I just want to make some images to add to fan fic.

I am going to go with the premise that KSP1 was retro-future esque


Hubris ... and asteroid mining gone awry let do several decades of collapse  / restructuring



I like when i applied a slight cartoon weight on the pictures of the Collapse (Kerlapse?).

My  attempts at Kerbals though, not as amazing. Without a plugging any LORA  i ended up with helmets almost any time i  used the word Kerbal. 


Now i am curating a select group of images of Kerbals without helmets to train a model and hopefully get better looking kerbals.

not quite finished but it is getting closer to the what i want. Finding more good pictures of kerbals without helmets will  help.
I have the aesthetic i want but the shape of the head is still of lol.



Look forward to following your story.