Copita – Train a style LoRA from a single image (SDXL / Anima)

Copita is a Windows app that trains a style LoRA from one image.How it worksGive it a picture drawn in the style you want. Copita writes you a prompt; you paste that prompt and your picture into an image-editing AI such as ChatGPT, and get back the same character in the same pose, rendered as a photograph instead of an illustration.Now you have two pictures that differ in exactly one thing.Copita trains a LoRA that can reproduce nothing but picture A, another that can reproduce nothing but picture B, and subtracts them. The character cancels. The pose cancels. The background cancels. What survives the subtraction is the linework, the shading, the palette, the brush feel — the style.One button runs the whole loop, captioning included.Three presets Stable — 200 steps, dim 8. For 12GB cards. About 50 minutes on an RTX 3060 Normal — 300 steps, dim 8. The balanced default Detailed — 400 steps, dim 16, alpha 8, colour emphasis on. Tens of minutes on an RTX 4080 Also in the box Concept sliders. Feed it a pair that differs only in line weight, body proportions or detail density, and you get a slider instead of a style. The "Bold LoRA for Anima" on my Civitai page was made this way A chibi template, for when the body proportions themselves are what you want to learn Per-module learning rates. Training self-attention only is what keeps a slider from repainting your picture A Fusion tab that merges 2 to 4 finished LoRAs into one cleaner result One thing to understand before you buyThe subtraction removes what is genuinely identical in both pictures — not what you personally consider "not style". If your reference has water droplets on the skin and the photoreal twin does not reproduce them, your style LoRA will add water droplets to everything. Speech bubbles, sweat marks, a signature in the corner: same story.Pick a clean reference — a simple bust-up, simple clothes, simple background — and you get a clean LoRA. The included guide covers this properly. It is the whole skill of the method.Requirements Windows 10 or 11 NVIDIA GPU, 12GB VRAM or more About 20GB free disk space. The installer builds its own Python environment and fetches sd-scripts. It does not touch your existing setup, your PATH, or your Forge / ComfyUI install Access to an image-editing AI to make the paired picture. Check that service's terms permit it One picture Notes Trained against anima_baseV10, so Anima style LoRAs carry across derivative models. SDXL supported Japanese and English UI The long-form documentation and my experiment write-ups are in Japanese. The app and the quickstart guide are fully English Made by Sakaki Io from studio Masakaki.Official X：https://x.com/studiomasakakiFANBOX：https://studiomasakaki.fanbox.cc/

Found on Devadex — the discovery index for independent software the big search engines bury. More from gumroad.

Report this listing