Drafted (YC P26) announced models trained specifically for residential architecture workflows. The launch appeared on Hacker News where the thread collected 24 points and 22 comments.
The models focus on floor-plan generation, code compliance checks, and material quantity estimates for single-family homes. Training data centers on U.S. residential building codes and standard construction practices rather than general image datasets.
What It Is and How It Works
Drafted fine-tunes diffusion and transformer architectures on paired architectural drawings and permit documents. Inputs include site dimensions, zoning constraints, and room counts. Outputs return vector floor plans plus basic 3D massing that comply with local setback and height rules.
The system separates structural layout prediction from code validation. A second model layer flags violations against IRC and local amendments before export to CAD formats.
Early Community Metrics
The Hacker News discussion reached 24 points with 22 comments in the first 48 hours. Commenters noted the narrow residential focus as a departure from general-purpose tools.
No public benchmark numbers for FID scores or compliance accuracy appear in the launch post. Early testers referenced internal validation sets drawn from 12 U.S. municipalities.
How to Try It
Access currently requires joining the Drafted waitlist through their site. The company plans to release a limited API for qualified architecture firms by the end of the YC batch.
No public Hugging Face weights or open weights are listed. Integration examples in the thread point to future Revit and ArchiCAD plugins.
Pros and Cons
- Narrow domain training reduces hallucinated room layouts common in general image models.
- Built-in code checks cut manual review time for permit submissions.
- Limited to residential projects; commercial or multifamily work falls outside current scope.
- No public performance numbers yet, making direct speed or accuracy comparisons difficult.
Alternatives and Comparisons
General models such as Stable Diffusion XL and Midjourney require heavy prompting and post-processing to produce code-compliant plans. Drafted embeds compliance logic directly.
| Feature | Drafted Residential | SDXL + Custom LoRA | Midjourney v6 |
|---|---|---|---|
| Code compliance | Built-in | Manual | None |
| Output format | Vector + 3D | Raster | Raster |
| Training domain | U.S. residential | General images | General images |
| Public benchmarks | None yet | Community sets | None |
Who Should Use This
Residential architects handling 1-3 story homes in jurisdictions with standard IRC adoption gain the clearest workflow improvement. Firms already using Revit or ArchiCAD can test the planned plugins first.
Practitioners working on multifamily, commercial, or international projects should continue with existing general tools until Drafted expands its dataset.
Bottom Line
Drafted supplies the first publicly discussed domain-specific models that combine layout generation with automated residential code validation.
The narrow focus and lack of open benchmarks position it as a specialized productivity layer rather than a replacement for broad creative tools. Future API releases will determine whether adoption spreads beyond early YC network users.
Specialized vertical models like these reduce the prompt engineering overhead that general image generators impose on technical fields.
Top comments (0)