lordspline/canvas-4b-design-model-proposal

Canvas-4B: Design-Native HTML Generation

This repository contains Capy Research's literature review through 1 August 2026 and a complete proposal for a narrow, approximately 4B-parameter multimodal model that supports:

1. screenshot → faithful single-file HTML reconstruction;
2. screenshot + brief → high-quality HTML redesign; and
3. brief → high-quality HTML creation.

Read the 30-page Revision 2 paper inline

The proposal recommends a Qwen3.5-4B-Base derivative with a shared native multimodal backbone, protected local visual tokens, mode-specific low-rank adapters, training-only structure heads, and render-grounded optimization. Revision 2 fixes the model-facing data contract to nullable image, nullable text, and complete HTML; specifies every existing-dataset conversion; and defines CanvasMix-1.3M as one million broad competence records plus an expandable 300K TasteCore. The internal noncommercial experiment records provenance for a later commercial rebuild rather than using commercial readiness as its admission gate.

Repository contents

• docs/canvas-4b-design-model-proposal.pdf — canonical paper.
• source/ — deterministic LaTeX/Markdown build sources and visual assets.
• research/sources.tsv — the 37-paper primary-source corpus reviewed.
• research/hf-dataset-audit-v2.json — the 1 August 2026 audit of candidate Hugging Face datasets.
• SHA256SUMS — checksums for every published artifact.

Reproducibility

From source/, install Pandoc, Latexmk, and LuaLaTeX, then run:

python3 render.py
python3 render.py --check

The canonical PDF SHA-256 is:

f772cd1c442b36f16225c8b4580ec5d6672b2f4161ee754094bd239a525961de

Rights and use

This is a research proposal, not a released model or training dataset. Every cited model, dataset, font, and upstream artifact retains its own terms. Revision 2 distinguishes the broad internal feasibility corpus from the later commercial corpus that would be rebuilt with commercial eligibility enforced.