General description
Turns any picture into coloured terminal ASCII art.
The whole frame is built from monospace characters . : - = + * # % @ on black, each tinted with the colour of the picture underneath.
Brightness comes from how dense the character is, and colour comes from the tint, just like a real terminal image converter.
Works on characters, landscapes, still life, animals and city scenes. At strength 0.8 you get a hybrid: the character stays drawn, with glyph texture over it and a background made of characters.
Trained on synthetic data only: plain base-Anima renders converted by a script, so no other artist's style is mixed in.
────────────────────────────────────────────────────────────
Anima
Base model: Anima (anima-base-v1.0)
Trigger: @asciiterm at the start of the prompt (required)
Style tag: ascii art at the end of the prompt. It more than doubles how much of the frame turns into characters.
LoRA strength: 1.0 for full ASCII, 0.8 for the hybrid look
Example prompt
@asciiterm, safe, masterpiece, best quality, score_7, 1girl, frieren, elf, pointy ears, white hair, twintails, green eyes, white capelet, flower field, blue sky, upper body, ascii art
Recommended settings
Sampler: DPM++ 2M SDE, scheduler simple, 25 steps, CFG 3.5 (3–4)
Size: 1216×1824 / 1824×1216 / 1472×1472 (≈2.2 MP, fits the Civitai limit)
Hires fix: off. Upscaling stretches the characters and the second pass blurs them, so generate at the large size directly.
The characters always come out the same pixel size, so a bigger frame fits more columns: at about 2.2 MP every glyph is crisp. 1 MP (832×1216) still works, but the characters look softer.
Negative prompt:
worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, watermark, signature, bad anatomy, bad hands, nsfw
Tips
Lit scenes read best: sunsets, lanterns, a lit office, a flower field. Very dark night scenes can come out too dark at the large size.
Bright daytime scene not converting fully? Add white background, bright, daylight, smooth shading to the negative.
Keep DPM++ 2M SDE: with plain Euler the effect drops sharply. CFG 7 and above adds nothing.
Keep "safe" in the prompt and "nsfw" in the negative. Base Anima leans towards NSFW on its own.
Don't add extra aesthetic tags like score_8 or aesthetic.
────────────────────────────────────────────────────────────
Training
120 synthetic images: plain base-Anima frames (50 known characters, 50 original characters, 20 scenes without people, mixed aspect ratios) turned into coloured ASCII by a script: contrast and saturation boost, Consolas Bold in a fixed 10×20 px cell, density ramp .:-=+*#%@, per-cell colour on black, crop to exact training buckets
Glyph size picked with an Anima VAE round-trip test: 7 px turns to mush, 9 px holds, 11–13 px is perfect but coarse
Captions: trigger + the source frame's content tags, nothing about the style
Anima LoRA, rank 32 / alpha 32, incl. AdaLN modulation · AdamW8bit, LR 1e-4, constant with 5% warmup · epoch 8 of 12 (960 steps), batch 1 · 1024 res, buckets 512–1536 · sigmoid timesteps, scale 1.3
Epoch 8 picked by comparing epochs 3–12 on scenes not in the dataset; settings come from a one-change-at-a-time study that measured how much of each image is a character grid