Merge pull request #6 from andrinr/diffable_render

feat: Differentiable Rendering / Scene fiting

Author: andrinr

.vscode/settings.json

@@ -5,9 +5,21 @@ Differentiable SDF primitives, transformations, and constraint system built with
 > [!WARNING]
 > The API is not stable. Expect breaking changes.
 
-![primitives](examples/ior.png)
+![primitives](examples/assets/primitives.png)
 
+---
+
+## Features
+
+- **SDF primitives** — sphere, box, capsule, cylinder, torus, and more
+- **Boolean ops** — union, intersection, subtraction with smooth blending
+- **Transforms** — translate, rotate, scale, mirror, repeat
+- **Raymarcher** — sphere-tracing renderer with materials, lighting, refraction, and anti-aliasing
+- **Differentiable rendering** — gradients flow through the full render pipeline via JAX
+- **Constraint system** — geometric constraints (distance, angle, coincident) with Riemannian gradient descent and Newton projection onto the constraint manifold
+- **JAX-native** — every scene is a pure function; `jit`, `grad`, and `vmap` work out of the box
 
+![primitives](examples/assets/constrained_optim.png)
 ---
 
 ## Development install
@@ -37,125 +49,12 @@ quarto preview    # serve locally at localhost:4321
 
 ---
 
-## SDF example
-
-Build a scene from primitives, boolean ops, and transforms:
-
-```python
-import jax.numpy as jnp
-from jaxcad.sdf import Sphere, Box, Capsule, Cylinder, Torus, Union, Translate
-
-sphere  = Translate(Sphere(radius=0.6),  offset=jnp.array([-1.0, 0.0, 0.0]))
-box     = Translate(Box(size=[0.7, 0.7, 1.0]), offset=jnp.array([0.0, 0.0, 0.8]))
-capsule = Translate(Capsule(radius=0.3, height=1.3), offset=jnp.array([1.0, 0.0, 0.0]))
-
-scene = Union((sphere, box, capsule), smoothness=0.1)
-
-# Evaluate the SDF at any point
-p = jnp.array([0.5, 0.0, 0.0])
-print(scene(p))  # signed distance from p to the surface
-```
-
-Every node is a pure JAX function — `jax.grad`, `jax.jit`, and `jax.vmap` work directly on the scene.
-
----
-
-## Rendering example
-
-Assign materials to primitives and render with the sphere-tracing raymarcher.
-`background_color`, `refract_steps`, and `ior` are all new in this release.
-
-```python
-import jax.numpy as jnp
-from jaxcad.render import raymarch, Material
-from jaxcad.sdf.primitives import Sphere
-from jaxcad.sdf.boolean import Union
-from jaxcad.sdf.transforms import Translate
-
-# Glass sphere (ior=1.5) in front of two coloured spheres
-glass = Sphere(
-    radius=1.0,
-    material=Material(color=[0.92, 0.97, 1.0], roughness=0.05, opacity=0.04, ior=1.5),
-)
-red   = Translate(Sphere(radius=0.65, material=Material(color=[0.93, 0.26, 0.22])),
-                  offset=jnp.array([-1.1, 0.5, -3.0]))
-green = Translate(Sphere(radius=0.65, material=Material(color=[0.05, 0.72, 0.50])),
-                  offset=jnp.array([ 1.1, -0.5, -3.0]))
-scene = Union((glass, red, green), smoothness=0.0)
-
-image = raymarch(
-    scene,
-    camera_pos=jnp.array([0.0, 0.5, 5.5]),
-    resolution=(400, 400),
-    background_color=jnp.array([0.07, 0.09, 0.16]),  # dark night sky
-    refract_steps=48,   # two-bounce Snell's-law refraction
-    aa_samples=2,
-)
-# image is a (400, 400, 3) float32 numpy array
-```
-
----
-
-## Constraint example — Riemannian gradient descent
-
-Move a point along a constraint manifold using Riemannian gradient descent: gradient steps stay on the tangent plane and a Newton projection snaps back to the manifold after each step.
-
-```python
-import jax
-import jax.numpy as jnp
-import optax
-
-from jaxcad.constraints import (
-    DistanceConstraint, Vector,
-    null_space, make_manifold_projection,
-)
-from jaxcad.extraction import extract_parameters
-
-# Constrain p to lie on the sphere |p| = 2
-anchor = Vector(jnp.array([0.0, 0.0, 0.0]))
-p      = Vector(jnp.array([2.0, 0.0, 0.0]), free=True, name="p")
-DistanceConstraint(anchor, p, distance=2.0)
-
-free_params, _, metadata = extract_parameters(p)
-target = jnp.array([1.0, 1.5, 0.0])
-
-def objective(params):
-    return jnp.sum((params["p"] - target) ** 2)
-
-value_and_grad = jax.value_and_grad(objective)
-
-def riemannian_grad(params):
-    """Project gradient onto the tangent plane at the current point."""
-    N = null_space(params, metadata)          # tangent-space basis (relinearized)
-    loss, g = value_and_grad(params)
-    return loss, N @ (g @ N)                  # Riemannian gradient
-
-# Riemannian GD: tangent-plane steps + manifold projection after each update
-optimizer = optax.chain(
-    optax.sgd(0.15),
-    make_manifold_projection(metadata),       # Newton snap-back onto |p|=2
-)
-
-params = free_params
-state  = optimizer.init(params)
-for step in range(20):
-    loss, g = riemannian_grad(params)
-    updates, state = optimizer.update(g, state, params)
-    params = optax.apply_updates(params, updates)
-
-print(params["p"])  # [1.109, 1.664, 0.] — optimal point on |p|=2 closest to target
-```
-
-`null_space` recomputes the constraint Jacobian at the current point each step, so the gradient is always projected onto the correct tangent plane. `make_manifold_projection` chains as a standard optax transform and works with any base optimizer.
-
----
-
 Inspired by [Fidget](https://www.mattkeeter.com/projects/fidget/) and [Inigo Quilez's distance functions](https://iquilezles.org/articles/distfunctions/).
 
 ---
 
-![primitives](examples/thingy.png)
+![primitives](examples/assets/thingy.png)
 
 ## License
 
-[GNU Affero General Public License v3.0](LICENSE) — free for open source use; commercial use requires a separate license. Contact the authors if you want to use jaxcad in a proprietary product.
+[Elastic License 2.0](LICENSE) — free for personal, research, and internal business use. Offering jaxcad as a hosted or managed service requires a commercial license. Contact [andrin.rehmann@simulation.science](mailto:andrin.rehmann@simulation.science) for commercial enquiries.

LICENSE

@@ -47,7 +47,13 @@ quartodoc:
         - extraction.extract_parameters
     - title: Functionalize
       contents:
-        - sdf.functionalize.functionalize
+        - functionalize.functionalize
+        - render.functionalize.functionalize_render
+    - title: Parametrization
+      contents:
+        - parametrization.to_normalized
+        - parametrization.from_normalized
+        - parametrization.compute_param_scales
     - title: Constraints — types
       contents:
         - constraints.types.base.Constraint
@@ -61,18 +67,12 @@ quartodoc:
         - constraints.solve.project_to_manifold
         - constraints.solve.constraint_residuals
         - constraints.solve.make_manifold_projection
-    - title: Construction
-      contents:
-        - construction.from_point.from_point
-        - construction.from_line.from_line
-        - construction.from_circle.from_circle
-        - construction.extrude.extrude
     - title: Geometry
       contents:
         - geometry.parameters.Vector
         - geometry.parameters.Scalar
     - title: Render
       contents:
-        - render.raymarch.raymarch
-        - render.raymarch.render_raymarched
+        - render.raymarch.render.raymarch
+        - render.raymarch.render.render_raymarched
         - render.marching_cubes.render_marching_cubes