Architecture · Hybrid Neural + Vector Synthesis

Decoupling stick-figure geometry from hand-lettering

Pure image diffusion models garble speech-bubble text. xkcd-artist masks dialogue out during L4 training, runs a 1.72 MB Int8 ONNX planner in WebAssembly, and composites crisp xkcd-script SVG ink.

xkcd-artist hybrid neural and vector synthesis architecture Architecture diagram showing offline masked XKCD training feeding the gnumanth/xkcd-artist model hub, which supplies Int8 ONNX weights to an in-browser WebAssembly planner and SVG ink compositor. OFFLINE L4 TRAINING & HUB IN-BROWSER WASM & VECTOR ENGINE L4 TRAIN INT8 WEIGHTS PROMPT TENSORS 5 HEADS PROPS + TTF VECTOR DATA Masked Corpus 226 PNGs · 891 scripts HUB gnumanth/xkcd-artist HF · SD LoRA + ONNX USER Prompt / WebMCP ?q= · modelContext TOK MD5 Tokenizer int64[1,36] · cast[1,2] WASM ONNX Int8 Planner 1.72 MB · 5 heads · 5.5ms INK SVG Compositor 4x wobble · leaders ASSET Stage & Font Pack 11 props · xkcd-script OUT XKCD Strip SVG · 2x PNG · alt LEGEND Focal planner Engine step Corpus / assets Model hub User / agent Neural path Hub weights Sync / build

01 · SPEECH-BUBBLE-MASKED TRAINING

Zero diffusion text hallucinations

Dialogue in the upper 36% of 226 XKCD panels is masked out before NVIDIA L4 fine-tuning alongside 891 ExplainXKCD scripts, teaching pure stick-figure and stage-prop ink geometry without garbled letters.

02 · IN-BROWSER ONNX WASM

5-head structural planner in 5.5 ms

  • Layout family (1–4 panels, chart, flowchart)
  • Stage prop & 12-DOF stick kinematics
  • Dynamic pen wobble, stroke weight & abstention

03 · VECTOR COMPOSITOR

Crisp xkcd-script & hover alt-text

Hand-wobbled #0A0A0A pen paths, head-crown leader lines, CC-licensed xkcd-script lettering, and Randall-style hover title text render at infinite SVG resolution.

xkcd-artist · hybrid neural + vector architecture gnumanth/xkcd-artist · 1.72 MB Int8 ONNX WASM