<h2 id="vintageanima">VintageAnima</h2>A full finetune of Anima on old school anime screenshots themselves.VintageAnima was trained on 92,308 anime screenshots from 91 films and series. Eleven of those titles are late 1970s, seventy are 1980s, and a handful of early-90s films are still drawn in that same tradition. These are the cels and the TV prints from that period: the ink, the paint, the film grain, the dust, the scratches, the faded color, and the character designs of the people who drew them.This is late-70s and 80s anime as it looked on a screen. The line, the palette, the proportions, and the dirt come from those frames. A clean painted cel is in there, so is the worn screening print. Modern "classic style" work is typically a new drawing emulating older screenshots with fake FX, and sometimes produce overly-clean results. VintageAnima was trained on only authentic, real screenshots for a faithful reproduction or transposing of modern characters into a classic style the way an older studio would have done it.<h3 id="prompt-it-the-way-it-was-trained">Prompt it the way it was trained</h3>Captions follow one order:<code>anime screenshot, year, decade, characters, series, description, hires or lowres, character designer, quality</code>Name the year and the series. Name the character designer when you know it. Describe hair, eyes, and clothes. If you leave the costume out, the model will dress the character for the series you named. A short pose sentence can sit in the description with the tags. About a fifth of the training captions were written that way.Quality is the print. <code>hq cel animation</code> is the clean painted cel. For the grungier screening copy, use the tags that were actually on those stills: <code>film grain</code>, <code>dust and scratches</code>, <code>faded colors</code>. <code>retro artstyle</code>, <code>traditional media</code>, and <code>thick outlines</code> pull the drawing back toward the period.Two tools build that caption for you:- Prompt studio (<a target="_blank" rel="ugc" href="https://zetaneko.github.io/vintage-anima/">https://zetaneko.github.io/vintage-anima/</a>) is the catalog. It has the 91 series, the cast, the designers, and the same caption order as training.- ComfyUI node (<a target="_blank" rel="ugc" href="https://github.com/zetaneko/vintageanime-comfyui">https://github.com/zetaneko/vintageanime-comfyui</a>) is <code>VintageAnima Prompt</code>. It writes the positive and the negative inside the graph. On a hires caption the negative takes the lowres-only quality tags. On a lowres caption it takes the hires-only tags.Note that an additional 20% of the dataset includes a blend of natural language captions to retain that style of prompting as did the original Anima model training.<h3 id="use-it-like-anima">Use it like Anima</h3>This file is the finetuned Anima diffusion model. Pair it with Anima's text encoder and VAE.- CLIP: <code>qwen_3_06b_base</code> (cosmos)- VAE: <code>qwen_image_vae</code>- Sample settings: 30 steps, CFG 3.5, sampler <code>er_sde</code>, scheduler <code>beta</code>It knows the cast it was trained on, from the late-70s heroines to the 80s city, fantasy and mecha shows. It will also shoot a character Anima already knows as one of those cels. The series in the prompt is the film stock. The clothes you write are the costume.<h3 id="closing-notes-creator-commentary">Closing Notes / Creator Commentary</h3>This model is an experiment that I think went very well relying on practically all automation with Grok Build from screenshot extraction, character detection, captioning and bucketing for num. repeats to ensure the dataset is diverse between artstyles, characters, genders etc. I chose this specific timeframe and limited anime because I wanted to validate how Anima would handle training on this kind of material at a small-scale before going big. One of the things I'll try to do next time is de-emphasize any character images detected from far in the frame that is zoomed in which can create some overly-simplified character stills sometimes.Would I like to do another round of training but with a much larger library of anime possibly going back to the 60s and up to the 90s? Yes, but I will wait and see if there is a more competent base model in the future before I invest money into a larger, full-scale training regime. Fundamentally, I want to stick to anime models which have a lot of existing knowledge specifically on Booru/anime scene concepts that most models like Krea2/Flux simply dont have, and finetunes don't capture enough anime concepts as well as something like Anima, Pony etc.Also, just a bit of a whine but I got a hand on NovelAI V5 and it can do some retro 80s styles nicely, but if you introduce any modern concepts or anything that isn't SFW it collapses and does some gross looking things. Partly why this model was created in the first place. Dont want to throw shade at NovelAI but it is just a niche/edge-case flaw with it.

VintageAnima (Very Old Anime Style)

Full finetune of Anima on 92,308 anime screenshots, 91 titles, weighted to the late 1970s and the 1980s. Same CLIP and VAE as Anima <code>qwen_3_06b_base</code>, <code>qwen_image_vae</code>). Samples: 30 steps, CFG 3.5, <code>er_sde</code>, <code>beta</code>.Prompt studio: <a target="_blank" rel="ugc" href="https://zetaneko.github.io/vintage-anima/">https://zetaneko.github.io/vintage-anima/</a>ComfyUI node: <a target="_blank" rel="ugc" href="https://github.com/zetaneko/vintageanime-comfyui">https://github.com/zetaneko/vintageanime-comfyui</a>

Files with SHA256: b3ded4185fdb1302a992b3897864b0103a0f539cb58efc47ce06f8e61c6814e5