{"version":"v1","site":{"name":"expectedwrong","url":"https://expectedwrong.com"},"links":{"collection":"https://expectedwrong.com/api/public/posts","rss":"https://expectedwrong.com/rss.xml","llms":"https://expectedwrong.com/llms.txt"},"post":{"slug":"gorilla-giraffe-akira-lora","title":"The Gorilla and the Giraffe Walk Into a Bank","subtitle":"Fine-tuning a LoRA on Akira and discovering that style transfer is basically just theft, but a really good kind.","url":"https://expectedwrong.com/gorilla-giraffe-akira-lora","api_url":"https://expectedwrong.com/api/public/posts/gorilla-giraffe-akira-lora","published_at":1723896000,"published_at_iso":"2024-08-17T12:00:00.000Z","updated_at":1771546355,"updated_at_iso":"2026-02-20T00:12:35.000Z","tags":["lora","image-generation","fine-tuning","akira","ai-art"],"excerpt":"Fine-tuning a LoRA on Akira and discovering that style transfer is basically just theft, but a really good kind.","meta_description":"Fine-tuning a LoRA on Akira and discovering that style transfer is basically just theft, but a really good kind.","reading_time_minutes":2,"word_count":257,"engagement":{"signals":0,"counterpoints":0},"body_markdown":"Been fine-tuning a cinematic LoRA.\n\nThe test prompt is a story about a gorilla and a giraffe who go on a bank robbing spree — which is, arguably, the correct prompt for evaluating any image model. If your model can't render armed primates in a heist scenario with a giraffe getaway driver, what are you even doing.\n\nThe original scenario image is fine. Competent. It has a gorilla, it has a giraffe, they appear to be up to something.\n\nThen you run the same prompt through the Akira model.\n\nAnd that's when it becomes a different kind of problem — the kind where you close your laptop, walk around the room, and come back to check if you actually saw what you think you saw. The heavy ink lines. The speed. The exact way neon reflects off wet pavement at 2am in a city that doesn't technically exist. The gorilla looks like it should be on a three-story wall mural in Neo-Tokyo.\n\nThis is the thing about style transfer that nobody talks about seriously enough: you're not really generating an image, you're excavating the visual logic of a human artist — the decisions Katsuhiro Otomo made in the eighties about how shadow works, how motion works, how dread works — and then repurposing that logic to render a giraffe holding a duffel bag full of cash.\n\nOtomo didn't consent to this. Neither did the gorilla.\n\nThe Akira LoRA is cracking, though. That's the word for it. Not good, not impressive — cracking. Which is its own category.","body_text":"Been fine-tuning a cinematic LoRA. The test prompt is a story about a gorilla and a giraffe who go on a bank robbing spree — which is, arguably, the correct prompt for evaluating any image model. If your model can't render armed primates in a heist scenario with a giraffe getaway driver, what are you even doing. The original scenario image is fine. Competent. It has a gorilla, it has a giraffe, they appear to be up to something. Then you run the same prompt through the Akira model. And that's when it becomes a different kind of problem — the kind where you close your laptop, walk around the room, and come back to check if you actually saw what you think you saw. The heavy ink lines. The speed. The exact way neon reflects off wet pavement at 2am in a city that doesn't technically exist. The gorilla looks like it should be on a three-story wall mural in Neo-Tokyo. This is the thing about style transfer that nobody talks about seriously enough: you're not really generating an image, you're excavating the visual logic of a human artist — the decisions Katsuhiro Otomo made in the eighties about how shadow works, how motion works, how dread works — and then repurposing that logic to render a giraffe holding a duffel bag full of cash. Otomo didn't consent to this. Neither did the gorilla. The Akira LoRA is cracking, though. That's the word for it. Not good, not impressive — cracking. Which is its own category.","hindsight":{"verdict":"persists","note":"LoRA aesthetic fine-tuning is still the gorilla-and-giraffe-in-a-heist-movie test. The Akira style works. The process hasn't changed. The prompts are still unhinged.","links":[],"at":1739980800,"at_iso":"2025-02-19T16:00:00.000Z"}}}