Fabivo / image-to-CAD research

From furniture photos to editable boards

Predict a structural drawing. Read its boards. Compile actual CAD panels.

Saved results / no live inference / no GPU calls

GPT Image 2.5 vs Fabivo m11

Two GPT image edits versus one direct fine-tuned drawing. Both real drawings enter the same Fabivo reader; no boards are manually repaired.

White staggered console: reference photograph, GPT Image 2.5 drawing and compiled CAD, and m11 drawing and compiled CAD.
White staggered console. GPT uses a saved rerun after the first provider response failed. Select the figure to view it at full size.

Illustrative archived cases, not a controlled benchmark. Prompts, resolution and budgets differ. GPT: Azure gpt-image-2.5-sunburst, low quality, frontal candidate 1, two-step-v5. m11: published rank64 3000-EMA, Turbo10, 768 pixels, seed 0. CAD defaults: 1200 mm width, 350 mm depth, 18 mm boards. Not measured dimensions or a safety-approved cutting plan.

All three cases and exact protocol

The ten requested references

All ten requested files have real saved m11 drawings and compiled documents. Six were generated in one bounded Modal batch; four reuse verified archived outputs. Example 04 repeats the staggered console with different encoding; it is not an independent unseen case.

Hugging Face refused ZeroGPU hosting for this account (HTTP 402). The owner then authorized one credit-limited Modal batch for the six missing predictions. This static gallery does not accept uploads, generate new images or run a persistent GPU server.