#battle & beasts

[1980sactionfigures]

Slasher Seahorse / Hunchback Camel - Battle Beasts (Hasbro)

[arcadebroke]

[effkaytales]

What may, in fact, have been my most cursed comic set yet

SG Lyzack's Beast Body has some... interesting quirks thanks to her adventures.

[sandmandaddy69]

[dawn-arts]

Could you Monsterify HUGE?

Yes! HUGE was kinda an enigma for me to design due to the bot's design being so unusually shaped.

However I think I've settled on this.

(The black comes from the armored wheels of the modern design)

Bonus image for size comparison with Sawblaze and a person:

These doodle designs are a nice break between Inktober posts so if anyone else wants a bot to be monsterfied, lmk!

[autoacafiles]

[dayzeebug]

We went op-shopping, and adopted some babies.

Beast, Trixie, Chief, Tarra, Hana, Cloud, and Stolas (a demon trapped in the body of an owl).

Thank you to whomever sent these kids into The Universe. And thank you to The Universe for sending them into my life.

[jimmyjamjemz]

Type Specialists for Battle Beasts.

Some of my favourite character designs that I've done!

[adam16bit]

Posted eight years ago today, it's Eledram Burst! It's a clear-ish blue elephant figure and it's pretty spectacular, also pretty obscure since these never really made it to the USA.

[tallus76-photography]

[askvectorprime]

Do Transformers celebrate any equivalent of Veteran's Day and Memorial Day? One where they honor soldiers who returned home and one where they honor those who gave their lives?

Dear Holiday Honored,

Sadly, war factors heavily into the Cybertronian experience, and commemorations of valor pepper our calendar like shrapnel. One such date was inspired by the sacrifice of Webslinger, who fought against the gluttonous and sedentary Builder, Rampage. Though Rampage crushed the arachnoid Protoformer with a datapad, Webslinger’s unwise sacrifice was an inspiration to many Predacons, and they hold an annual day of remembrance in his honor.

[1980sactionfigures]

Brain Mouse / Fly Sailor - Battle Beasts (Hasbro)

[rjmbaboonbooks]

Daily Comic Journal: July 30, 2015: “Uncovering Money.”

Here’s just a small sample of what I found:

[sandmandaddy69]

[deepdreamnights]

This Toy Does Not Exist: ManBeasts of Xor

Short lived but memorable. The regrettably sculpted and named “SteelHog” sank the whole premise, sadly.

An experiment with image prompting in Midjourney. Boring technical stuff about how AI tokens work below.

Prompt:  a battle beast action figure from Hasbro (1987)

For these, most prompts were a variant of the above line, with 1-4 Battle Beast toy photos used as image prompts. An example:

These were an experiment with image-prompting, mainly, seeing if I could get a highly accurate result for an obscure concept if backed up with image prompt support. The answer, not unexpectedly, is “no.” The ManBeests are clearly of a similar genre, but they aren’t remotely close to the trademark aesthetics of the line. Why so?

Say it with me “AI often doesn’t work how we assume it does.”

In the simplest terms, Midjourney, SD, Dall-E2, etc, all use a process by which two AI bounce JPEG compression noise back and forth until it looks like something. One AI is based on object recognition, the other primarily in image generation. The latter modifies the JPG noise with its series of learned patterns (more on that later) and the former tells it to do it over if it’s estimation of the “match” of image to prompt is worse than it was previously. Once they reach consensus, you’ve got a picture.

The process is more complicated than that in practice. There’s other AI processes looking for things the programmers don’t want, like porn and gore in the case of Midjourney, and correct against it. There’s general processes that influence composition. AIs with an ‘overpaint’ feature let you substitute an existing image for the random JPG noise as a starting point, etc.

The information the AI used to refine this JPG noise is gathered from the dataset, largely the LIAON sets scraped from the web. This information is not stored as pixel data. Instead, the AI breaks down the images into patterns based on similarities with other images sharing common tags. These “tokens” are what the AI builds its images out of.

When you type a prompt into the AI, that sentence is broken down into the associated tokens, and then the AIs get to work building the new image from that “vocabulary” of concepts. Essentially, the AI looks at a bit of JPG noise, goes “if this part looks like this, what’s the part around to it likely to look like, based on the weighted models of similar items I’ve seen in the past?” and then fills that in, repeating as its checked and re-checked by the image-assessment half of its function.

The use of abstract patterns rather than raw pixel data is why when you see the AI recreate a jibberish signature or watermark, it’s in generally the correct spot. Were the AI designed to collage bits of existing images together, as is often the assumption, you’d get chunks of signatures in random places and watermarks would be broken into randomly distributed chunks.

Instead, the pseudo-signatures tend to be in the right place. Because the AI learns from what it observes, and it’s learned that in a lot of pictures, there’s a bit of writing in the lower right-hand corner, and that its presence and prominence varies based on other tokens in play. It’s learned this the same way it learned where nostrils go.

So why is it that even with direct name and image prompting, the results we’re getting here are so... new?

Midjourney’s image prompting isn’t the overpainting process I mentioned above. Instead, the image is examined to see what tokens it recognizes, and it prompts for those as if it were part of the text prompt. (This does not add to the dataset, that would be a different process)

The thing is, no human laid down the token language. The AIs build them as they need them, and they form new tokens based on associations humans would not recognize, much less intend.

In a case where someone was able to back-engineer part of the token language of one of these image-AIs, they found that it had invented a new “word” that was associated with flying animals. A general catch-all term that grabbed all flying animals regardless of species, (funnily enough, very similar to the way “bird” is applied in the Bible.). This new term would be introduced into prompts as they were converted from English into the token language, based on what the text-parser deemed appropriate.

So it might know Battle Beasts, but doesn’t have a lot there. But it also has everything BB is associated with, even if its not explicitly called out in the prompt or images. I expect that either through the dataset, or just through aesthetic observation, it’s connected the concept of “Transformer” and the concept of “Battle Beasts” given the flourishes, but everything from the general concept of an action figure to the general idea of an animal is getting involved there.

In essence, no prompt is ever “pure”, and every generation involves far more influences than you’d imagine. Even taking every effort to narrow down the variables, the results were still unexpected.