mcpField
M
multiview-fit-loop

multiview-fit-loop

github.com/roble3Creative & media128838 installsUpdated 2w ago
Repository

About multiview-fit-loop

multiview-fit-loop is a skill for iteratively rendering, comparing, and adjusting 3D models against templates or originals. It measures silhouette and bounding-box differences across views and checks whether the supplied constraints are mutually consistent.

Key features

  1. 01Render flat silhouettes for each view
  2. 02Extract template object masks
  3. 03Compare bounds, centroids, and silhouettes
  4. 04Adjust models and render again
  5. 05Check consistency across view constraints

Use cases

  1. 01Improve model fit to orthographic templates
  2. 02Find size or center offsets between views
  3. 03Handle contradictory template constraints

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/roble3/cc-blender-skill --skill multiview-fit-loop

After installing, type/multiview-fit-loopin Claude Code to confirm it loaded.

Bundled filesscripts/multiview_fit_report.py

Details

From the project README

Multiview Fit Loop

This skill closes the missing loop: render → compare → adjust → render again. It is mandatory when a user says the model still does not fit the templates/originals.

Required loop

  1. Render flat, material-independent silhouettes for every available template view: front, side, back, top.
  2. Extract the template object mask, excluding labels, cyan guides, and background.
  3. Compare template vs render per view:
    • bbox center and size
    • centroid drift
    • silhouette coverage/IoU where modality is valid
    • visual overlay
  4. Convert measured deltas into model/camera/recipe adjustments.
  5. Rebuild or transform the model.
  6. Repeat until all hard gates pass or document the remaining conflict.

Constraint inconsistency gate

Before forcing adjustments, check whether the supplied orthographic templates are mutually consistent. A single rigid 3D model cannot simultaneously satisfy contradictory physical ratios, for example if side view says total depth is 0.39 of height but top view says depth is 0.98 of width. When this occurs, stop claiming final fit, write a conflict report, and create separate variants or ask which view is canonical.

Hard gates

  • all required views have validation reports and overlays;
  • bbox center drift <= 1.5% of image width;
  • bbox size drift <= 3% for front, <= 5% for side/top/back first-pass depth;
  • structural part count exact;
  • texture UV regions still valid after geometry changes.

View mask rule

For annotated wireframes, choose a template mask mode that isolates the intended construction/object lines and excludes guide colors, labels, captions, and background annotations. For Blender validation renders, use a flat white silhouette on black, not beauty renders with glow/context elements.

Scripts

  • scripts/multiview_fit_report.py compares view pairs and writes JSON + overlays.

Adjustment rule

Prefer changing recipe parameters or source geometry over camera scale tricks. Camera scale may be used only after model dimensions are correct.

Comments

Markdown supported