From-scratch 128x128 rectified-flow model draws newsprint faces on CPU in 0.27s, MIT licensed
rustydustyshckleford · reddit · 2026-11-02
grinpress v3 is a tiny from-scratch 128x128 rectified-flow model that draws newsprint-style faces with text beside them, MIT-licensed with a live CPU demo on Hugging Face.
- Character-level text conditioning: 7 expressions, colors, paper styles, layouts, up to 14 characters with mixed case and punctuation
- v3 distills a 50-step teacher into 2- and 4-step students, exported to ONNX with dynamic INT8 quantization
- CPU-only inference: 0.27s at 2 steps / 0.54s at 4 steps on an 8-vCPU Intel Xeon VM — no GPU needed
- Limits: narrow domain, long phrases get cramped, occasional mangled characters
More from Multimodal
- AI-generated "next Batman" portraits are gloriously ridiculous — egeberkina · 2026-11-02
- Lightx2v releases RL-refined Qwen-Image-2.1 Lightning v2: 4-step generation and editing — linoy_tsaban · 2026-11-02
- Dev releases ComfyUI Universal Regional Prompter: paint regions inside the node — eimissuper · 2026-11-02
- Dev vibe-codes a path-traced 3D editor with Seedance rendering inside Runway App Studio — jongranskog · 2026-11-02
- RTX 5090 Generates 15s Videos in 2 Minutes: Minimax H3 Speed-Tuned ComfyUI Settings Shared — Free_Pressure8623 · 2026-11-02
- Open-source insect sticker library ships with agent-run generate-and-review pipeline — AlchainHust · 2026-11-02