


Side-by-side AI art comparison contrasting 2026 GPT-2
By @CaylaCatz ·
16 likes · 17 comments
Then and Now : Same prompt @monkeydude69 and I were chatting about how much AI imaging has changed. We were talking about using our old prompt from way back then and seeing what it makes now. So I did. This is the prompt: Pensive duchess sitting on a settee, long flowing cobalt hair, periwinkle detailed gauze ballgown trimmed in lace, ornate stone castle ballroom, golden stained glass windows, regency period, intricately detailed, octane render, art nouveau, anime
Tags
- ai art comparison
- stable diffusion 1.5
- gpt 2 generative ai
- generative ai art evolution
- fantasy princess illustration
- royal character art
- ai image generation progress
- anime ai art
- realistic ai render
- fantasy castle interior
- vintage ballgown art
- ai model benchmark
- side by side art comparison
- digital generative art
- ai rendering
- anime princess
About this image
This side-by-side generative AI art comparison showcases output differences between newer 2026 GPT-2 image generation and 4-year-old Stable Diffusion 1.5 models, featuring two fantasy royal character portraits. The left panel displays a soft, painterly anime-style blue-haired princess in a ruffled lace blue ballgown in an ornate cathedral-style castle throne room, while the right panel features a semi-realistic teal-haired queen in a flowing cream gown seated on a gilded purple upholstered armchair. The visual comparison highlights shifts in AI art rendering quality, style priorities, and generative model capabilities over years of development.
This ethereal digital anime fantasy illustration features a thoughtful royal princess with long flowing cobalt blue hair, adorned with a jeweled gold tiara and delicate floral hair accessories. She wears an elaborate layered lavender and royal blue ballgown trimmed with lace and floral details, resting on a tufted velvet sofa inside a grand sunlit gothic castle chamber, where warm golden light filters through tall stained glass windows. The soft, serene mood and rich cool blue and warm gold color palette creates a dreamy, romantic fairytale aesthetic.
This painterly fantasy concept art features a regal fictional royal woman with long vibrant teal hair wearing an intricate golden crown and elegant off-the-shoulder lace cream gown. She sits thoughtfully on an ornate baroque armchair with carved gilded wood framing and lavender purple upholstery, bathed in warm golden sunlight filtering through tall paned palace windows. The mood is quiet and stately, with a soft, rich color palette blending warm gold, cream, and cool teal tones.
Comments (17)
yeah, four years ago it was very different. Deformities was common. Eyes not matching was very common. Hands with 6 fingers, 3 arms, etc. I don't know why I didn't give up. Midjourney is based on that old way. You make 4 pictures and they are all bad -- you choose the best and use that for your start and evolve 4 more pictures -- these look a lot better but still not good -- again choose the best one and evolve. Now I write a prompt and generate. Most times I like the first one unless I change my idea
@justrob ·
@CaylaCatz i wonder if the model evolves. I always thought that flux 1.1 pro was .. flus 1.1 pro - but then i took a very old prompt and the same seed, and submitted it, and it gave a different image!!!
Great comparison! AI as eveodle super quick alsthough i also liek some of the old iamges. And Stable Diffsion 1.5 images are offten great to evolve them with models liek Realvislighing.
@justrob ·
its very good. Wait - do you think results change from day to day?
@justrob ·
stunning work - i find these types of comparisons to be wicked cool
@legion ·
Yes, I remember. It's a good thing technology has advanced so quickly.
@legion ·
There is a big difference.
Nice. Love the dresses
Overall improvement change is whenever the model comes out. Results change not from day to day, but generation to generation due to automatic random seed changes. Models like GPT2 don't let you see the seed number any more or set the seed number. I think it's too bad because when I'm comparing a prompt in English to the same prompt in another language, it's better for the comparison if it is the same seed. But it's a trend. Almost all the newer models no longer let you see the seed number
I've wondered that too. The model creator is the one that actually generates the image. The platforms are simply user interfaces that connect you to those models. So if it's generated at Google's place, for instance, that means Google can do ongoing updates. GPT 2 had a real problem with artifacting when it first came out but it seems to be a lot better now. Not gone but definitely improved.
Sign in to join the conversation.
More from @CaylaCatz
See all →
VideoFolk art title card for Brief Musical Paws featuring a decorated lute and traditional floral border motifs
Whimsical Wonderful Wednesdays fantasy comic of characters enjoying a cozy tea party with Cookie Monster and Cheshire Cat
Anthropomorphic cat and rabbit explorers flee a fiery horned demon emerging from crumbling ancient ruins
Anthropomorphic cat bard singing while playing a lute, wearing a decorated folk hat and tunic
Elder magical baker holds a mooncake beside an iridescent spirit dragon in a moonlit workshop
Stitched black cat doll holds a glowing lantern in a patchwork enchanted autumn forest
More like this
See all →Made on BudgetPixel
The models and tools behind posts like this one.
- SeeDream 5.0 Pro
- MAI Image 2.7
- MiniMax H3 Image
- AI Image Generator
- AI Anime Character Generator
- AI Fantasy Character Generator
Guide: NightCafe vs BudgetPixel (2026): Which Is Better for AI Artists?