Just a compare article

The LoRA's i've tested can be found here: https://civitai.com/models/360281/y2k-mad-weirdcore-dreamcore

You can get compare grid without prompts and comments in attachments ->

photo, creature, disturbing, depressing, melancholic

weirdcore photo Alien-themed anthropomorphic biomechanical android monster alien mushroom, . Extraterrestrial . 35mm photograph

photo Alien-themed anthropomorphic biomechanical android monster alien mushroom, Extraterrestrial, 35mm photograph

photo Alien-themed anthropomorphic biomechanical android (monster alien:2) mushroom, Extraterrestrial, 35mm photograph

I see here and later - [extract] version is very affected by 'mushroom' theme. If mushroom word is appeared in prompt, you can't just remove it, even with (:xx) weight modyfiers. Thats bad

Here's what's going on when prompt is 'normal' - no weird characters, many 'detailed' tags

weirdcore, A angelic mythical lion with spreading arms with glowing intricate transparent wings, symmetrical hyperdetailed texture, pearl filigree, concept

y2kmad, weirdcore, A angelic mythical lion with spreading arms with glowing intricate transparent wings, symmetrical hyperdetailed texture, pearl filigree, concept art

weirdcore, y2kmad, DnD character concept art, rpg character art, big round staring eyes, swamp, amphibian woman, swamp queen, seductive swamp witch, green skin, a slimy girl, tongue hanging, long thin tongue, webbed feet, prehensile toes, realistic, high quality, detailed, staring round eyes

weirdcore, y2kmad, house on the green hills

As i can see here - [finetune extract] is overfitted and making images almost identical to training set. Thats bad ;(

weirdcore, y2kmad, liminal space

an image with a surreal, distorted office setting featuring cracked paint on the walls and a human silhouette with many eyes

a weirdcore composition that explores the theme of a distorted suburban environment with a focus on cracked paint, many eyes, and a human silhouette in an unsettling setting

Also here i see that [finetune extract] tends to make 'similar' images even when prompt is different, thats also bad

a weirdcore scene with a liminal space incorporating elements of suburbian life, creepy mushrooms, and a pool surrounded by a dirty, lo-fi atmosphere

an eerie image set in a pool area with a liminal space vibe, incorporating elements of horror, strange mushrooms, and a distorted atmosphere

an image that combines horror and absurdity, featuring a collage of strange elements like distorted faces, cracked paint textures, and surreal office equipmen

a collage image filled with y2k symbols, featuring a cartoon girl standing on a lawn surrounded by toys and typical suburban houses. Include eye-catching elements such as Hello Kitty, rainbows, red mushrooms with eyes, and angel-winged eyes, capturing the y2k nostalgia and weirdcore aesthetics

here^ we can see, that [0.0.1 mix] making best image - all the image is stylized, no 'real' humans or other elements, good work with colors, composition. [0.0.1 extract] is little messy and overloaded. simple [0.0.1] and [0.0.2] is ok, but the appearence of realistic person is the thing that make me feel, that model works wrong

a surreal scene where a woman sits on the edge of a stone ledge, gazing at the sky adorned with a massive eye peering out from behind a cloud, blending elements of human presence with the bizarre and weirdcore

a disturbing image featuring a hybrid creature with a creepy baby head, teeth, and enormous eyes fused onto the body of a fly, embodying the weirdcore aesthetica disturbing image featuring a hybrid creature with a creepy baby head, teeth, and enormous eyes fused onto the body of a fly, embodying the weirdcore aesthetic

[Extract] version just making trash. [0.0.1 mix] making good image, just the way i want it to work, but it feel slightly overfitting here, because generated image is similar to the training dataset way TOO much :) [0.0.1] and [0.0.2] works ok, but not really 'weirdcore'. Simply weird

Verdict

At first i thought, that finetune extract must be the best one, because of "full" training, but as I see here it's the worst one (in my opinion), so the full training was waste of time :D

My favorite one is [0.0.1 mix] and [0.01]

[0.0.2] is making slightly "fryed" results - the image is too messy, too detailed

BUT! [0.0.1 extract] is the most stable as i can see. It has the most stable geometry/topology and the least messy image