March 12, 2026
Fonts, fights, and 1982 flashbacks
High fidelity font synthesis for CJK languages
AI revives a classic; comments split between 1982 nostalgia and “will this be useful”
TLDR: An AI tool revives the zi2zi idea to auto-generate CJK characters in new styles, promising faster font-making. Comments split between nostalgic history dives and a blunt debate over practicality, with the author pushing “usable in the real world” while skeptics ask if it can truly handle full, consistent font sets.
A new tool called zi2zi‑JiT promises AI‑assisted fonts for Chinese/Japanese/Korean characters, taking one character and a style sample to whip up the same character in your chosen look. It’s trained on hundreds of fonts and comes in two sizes, with quick fine‑tuning. But the real show is the comments: one early take calls it a sequel to the original zi2zi, now with a modern transformer backbone, while others immediately time‑travel to Cangjie (1982) and even a possible DOS port, and someone drops a throwback METAPOST paper. It’s HN archaeology hour and everyone brought receipts.
Then the author parachutes in: “Hi author here”, reviving zi2zi to finally make font gen practical. That lights the fuse: skeptics ask if this will help designers actually ship complete CJK fonts (thousands of glyphs!) or if it’s another flashy demo. Supporters cheer that fine‑tuning can take under an hour and the base model claims ~4GB VRAM for batch 16, while jokesters quip “H100? brb calling NVIDIA” and “2,000 epochs? My laptop just fainted.” There are earnest questions—can it handle weird kanji, messy handwriting, or keep style consistent across a full set?—and meme‑y ones like “Will it do grandma’s grocery list font.” Metrics look “good‑ish,” but the vibe is practicality vs. hype. If this really makes CJK font creation less soul‑crushing, commenters agree: it’s a big deal—if it actually delivers.
Key Points
- •zi2zi-JiT extends the JiT architecture for Chinese font style transfer with content and style encoders and multi-source in-context conditioning.
- •Two variants (JiT-B/16 and JiT-L/16) were trained for 2,000 epochs on 400+ fonts (300k+ images), with a per-font cap of 800 characters.
- •Evaluation over 2,400 pairs reports FID, SSIM, LPIPS, and L1 metrics for both variants, following the FontDiffuser protocol.
- •Users can set up the environment with conda/pip, download pretrained checkpoints from Google Drive, and generate datasets from fonts or glyph images.
- •LoRA fine-tuning is supported on a single GPU; a single font can be fine-tuned in under one hour on an H100, with example hyperparameters provided.