Hi @Ameseth (and @ all others)!
While I lack the sophisticated knowledge to fully understand [AI] (and always end up eagerly following links to “a new and very simple method to… [AI]” just to be confronted with crazy algorithm-math, jokingly thinking that maybe “AI is writing papers about AI, and ML/data scientists don’t want to admit they also don’t understand these papers”) - I’d say I made up for that, in part, with “excessive trial-and-error” during “lockdowns”, and especially utilizing the wonderful AI named CLIP (the “old” CLIP released now just over a year ago) “running local”.
I’ve since:
- Had an epic brainflick-moment when I realized that “all that hate about gender stereotypical bias about AI on social media” is, in fact, a problem of the human-NLP side of things; namely, the English language used in prompts, which naturally doesn’t assign “gender” to all-and-every noun. And that I, in fact, could use RuDALLE - AFAIK, a version (“fine-tune” [?]) of CLIP that stands out for taking Russian language prompts to generate SO MUCH more precise, highly differentiated output… By working on “my” side, the NLP prompt, with regard to “switching my language for input / prompts to a more complex one”. (Hold on, this is related to your work!). So, with a car being naturally “female” in RU, by changing the preceding adjective to wrongfully addressing “car” as neutrum or maskulinum (works the same in RU as in DE - high five, fellow German, because this is near-impossible to explain to a native monolingual EN speaker - I’ve tried…) - e.g. “huebscher Auto”, “huebsches Auto”, “huebsche Auto” - (beautiful car) to produce 1. an aggressive-looking sports car, 2. a normal car (for RU “car” = fem.), 3. a small european looking car.
This - technically - should also apply for changing how gentle / aggressive / neutral your cat looks.
Wait, how does that apply to the English DALL-E?
- I came across this when feeding CLIP an image of a friend on a bicycle in 2021 (in gradient ascent, so basically “getting back a text “opinion” for what CLIP “sees” in the image”: Instead of the typically-spot-on astonishing “nailed it!” response after just a few hundred iterations - CLIP went on a “racist” anti-German slur. Which to me - knowing AI doesn’t have emotions or intentions - sent me into a fit of laughing up to the point of being in tears for this top-class cabaret-style comedy impression that was a) unexpected, b) still understandable “how [the AI] came to the conclusion, albeit derailed” and c) thus generated maximum entertainment. I attached a very much excellent example of this; a photo of my “Alexa” I’ve given to CLIP. And then photoshopped to remove the “typographic attack vulnerability exploitation”, i.e. the German text. The difference is… Exactly as mentioned in the prior example.
PS: I am sorry this image “lacks professional language” with referring to the typographic issue as “CLIP’s OCD reading bias” - it was made and tailored to a young lady who deeply enjoys receiving images exactly THIS way, and I didn’t save another (without annotations) - my bad, and apologies to anybody who might find this “inappropriate” - that was neither the AI’s NOR my intent! I am adding this as I believe it’s a very useful background info with regard to the CLIP’s tokenizer / generated ‘opinion’ tokens as they still seem to vaguely apply for DALL-E 2 as via my trial-and-error (see below):
-
So I had to know what happened and found out: It was related to the abundance of German language text seen on the bike frame (a cargo bike), and this top-of-the-tips comedy event was known as a “typographic attack vulnerability”. And also, CLIP’s dataset is heavily weighted towards “English”, but is in no way limited to English - CLIP (as in, “the original”, as released “free to the public” by OpenAI [thank you so much, I owe you my Covid-maintained excellent mental health, not even kidding!]), knows a multitude of different languages - including quite a bit of German, and Russian, and…
-
Subsequently, I realized that CLIP “naturally” creates not on German longwords, but any-and-all languages longwords. “Probabantenna”, “particleweaving spiderrollercoaster”, and “artificialintelligence trippyeyes library” include some of “actual, unaltered, as-is” best_of returned by CLIP models for “looking at an image” (fun fact: The last mentioned was “CLIP looking at a screenshot of the AI’s activation atlas” 
).
-
These “opinions”, especially but not limited to “when feeding the init image the ‘opinion’ was based on” work really, exceptionally well for creating a desired output. Even if it’s a desired output as given by the AI itself, it’s absolutely marvellous.
More directly related rather than background information, with regard to your project:
While “bat” in German is “Fledermaus”, so “one word”, in Russian it is, literally translated: “Flying mouse”, with the space; whereas Fleder+NewWord doesn’t necessarily make as sense, but let’s put that to the test. First, though: using RuDALLE, I had really great results with “the old CLIP” for making “Flying Rats”:
[edit: I had to smash this together because “not allowed for new users” to post more than one attachment… Also contains stuff mentioned further below]
I already noticed that together with astonishingly amazing quality AND overall coherence, DALL-E 2 also doesn’t seem quite so “rigid” with regard to concepts that have very few (if any) actual “representations” in real life; such as: “soul”, or, of course: “AI”. which, in the “old CLIP”, basically was represented by “red, monocular vision”. Very clearly a hollywood-inspired stereotype - and no longer present as such a “rigid concept” now, meaning you can no longer count on receiving “a red Terminator-style monocular vision” just by mentioning “AI” in the prompt; so “things might be more complex now”, and the hundreds, likely even many thousands of things I made with CLIP - cannot just be applied to DALL-E to receive “the same, but better”. It’s, in fact, a whole new quality, a whole new level [of awesome].
So let’s put all of the above knowledge to the test with DALL-E 2! 
1.: летучая мышь - RU for: “flying mouse”. Result: #dalle generates mouse. Just mouse. Conclusion: Initial evidence points towards “this AI know multiple languages again, too!”
2.: летучаямышь - the same, just as a “longword”. Interesting result: It’s creates a flying-[stripped-input]. All are “animals”, two examples:
[edit: I had to smash this together because “not allowed for new users” to post more than one attachment… see above]
-
Fledermaus - German for “bat”. Generates… you guessed it: Bats! Perfect bats! One example:
Iris × DALL·E | Fledermaus
-
Bold move: “Fledermaus but as a Katze” (bat but as a cat): Not sure what’s going on here, but it looks like this could be fun!
[Edit: “Sorry, new users can only add two links to posts”]. Alas, I will have to post this later, albeit the links being to openAI’s domain; I will once I know I won’t be tripping a spam filter and be classified as “an AI, not a human”. Sorry! 
-
Flederkatze; a wingedcatbat: The AI appears a bit confused, but with interesting results:
[Edit: “Sorry, new users can only add two links to posts”]
-
FlederKatze; a wingedcatbat with green cateyes and an adorable kittyface and a fluffy tail.
Now AI & I are both confused, but I - personally - totally wanna pet that fluffy batcat’s belly!
[Edit: “Sorry, new users can only add two links to posts”]
-
Finally, copying parts of your prompt for “style hints”:
FlederKatze; a flying wingedcatbat with batwingfrontpaws, green eyes and an adorable kittyface and a fluffy tail. photorealistic, professional digital art
[Edit: “Sorry, new users can only add two links to posts”]
Those were the best two out of a one-shot attempt [4 generations], with the others having the same issue as you reported - wings on the back.
However, with one having “front-pawed wings”, this is “going in the right direction” - albeit needs more work to refine the prompt, including credits for the attempt, so I’m gonna stop here. 
I am sorry if this is “too verbose” and a wall-of-text, but I genuinely hope that being in-context, it will serve to be helpful for your successful creation of art-as-intended!
Wishing you good luck, success, and most of all - a lot of fun! 
Reese (Iris)
