Image quality degrades significantly when downscaling by >2×

Summary

Images downscaled by more than 2× display severe visual artifacts including pixelation, jagged edges, and loss of fine details like text and thin lines.

Technical Cause

When a transform has scale < 1.0 (downscaling), the renderer inverts it to map from output space back to source image space. The x_advance and y_advance vectors are calculated from this inverted transform to determine how far to step in the source image for each output pixel.

Example with scale(0.125) (8× downscale):

1. User transform: scale(0.125) — render 800px image at 100px width
2. Internal transform: scale(0.125).inverse() = scale(8.0) — maps output→source
3. Advances computed: x_advance = (8.0, 0.0), y_advance = (0.0, 8.0)
4. During rasterization:
   • The SIMD rasterizer processes 4 rows at once, calculating positions as [0, 1, 2, 3] × y_advance + start_pos
   • For output starting at position (10, 5) → source position (80, 40)
   • Four SIMD lanes sample at y-positions spaced 8 pixels apart: [40, 48, 56, 64]
   • Bilinear filter applies fixed offsets [-0.5, 0.5], sampling 2 pixels at each position
   • Result: Samples only 8 pixels total; should ideally sample 32 pixels (4 rows × 8 pixels per row) — 75% of data skipped

The x_advance and y_advance correctly determine where to sample based on the scale factor, but the sampling kernel remains fixed at 2×2 (bilinear) or 4×4 (bicubic) pixels regardless of how large the advance steps are. For proper 8× downscaling, each output pixel should ideally average an 8×8 block (64 pixels), but instead samples only 4-16 pixels, skipping 60-48 pixels respectively and causing severe aliasing and loss of fine details.

Solution

• Add progressive downsampling for better image quality at small scales

Start by tracing the renderer's inverted downscaling transform through the SIMD rasterizer, focusing on x_advance, y_advance, and the fixed bilinear or bicubic sampling kernels described in the issue. Reproduce the scale(0.125) case and compare its sampled area with the expected downsampled region. Done means progressive downsampling improves detail and reduces aliasing at small scales without regressing other transforms.