Render time is a budget, not a virtue. A PBR interior can be honest at 12 minutes or honest at 90; the extra hour is often 8K grout, 2000 samples, and six area lights that barely contribute. These tips apply to Blender Cycles, V-Ray, Corona, and similar path tracers. Real-time engines have a different budget (frame rate); this article is for stills and offline frames.

Lock a preview recipe before the beauty

Work at 25–50% resolution with a denoiser on until cameras and materials are approved. Only then raise pixel size and samples. Artists who “always render 4K to see the textures” spend the day waiting and still miss a wrong color space. Keep a lookdev HDRI and a chrome/gray ball; lighting notes are in HDRI and PBR lighting.

Samples should follow noise, not a forum number. An overcast exterior often cleans up earlier than a dark interior with many glossy floors. Use adaptive sampling if the engine has it. Denoising is normal in 2025; pair it with enough samples that highlights on metal do not crawl. Dedicated article: denoising in PBR renders.

Textures: resolution, count, and format

PBR packs are heavy because they are four or five images, not one. A hero tabletop can stay 4K. A corridor wall that is 15% of the frame can be 2K or 1K. Duplicate 8K copies of the same concrete on every floor plate will stall GPU renders via VRAM, not via “quality”. Guide: 1K, 2K, 4K, 8K.

Pack ORM or drop unused height maps in the final scene. Prefer PNG masters and let the engine compress; do not ship 32-bit EXR albedo into a still that never exceeds 0–1. Format choices: EXR, PNG, JPEG. Packing: PBR texture packing. Free maps at several sizes: Locker3d textures.

Lights, GI, and caustics you do not need

Every shadow-casting light is a multiplier. One HDRI plus two windows beats a chandelier of 24 spots. Turn off caustics unless the shot is a glass product. Clamp fireflies instead of doubling samples if a small glossy highlight is the only noise source. Portal lights (or equivalent) help interiors more than another 500 samples of empty corners.

If the scene is huge, instances and proxies matter more than texture format. Distant trees do not need 4K bark. For games, LODs and streaming are a different article: LOD and texture streaming.

CPU, GPU, and memory

GPU is usually faster for PBR stills until you run out of VRAM. Then the card pages or the render fails, which is slower than CPU. Shrink textures, instances, and displacement before you “switch to CPU because it is more accurate”—it rarely is for standard metal/rough materials. Comparison: GPU vs CPU rendering for PBR.

What a render actually computes (and why samples exist) is in what is a render. Spanish overview: qué es un render.

A short checklist before you walk away

  • Test camera at half res with denoise.
  • Hero materials 2K–4K; background 1K–2K.
  • Kill unused maps and unused lights.
  • Confirm color space so you do not “fix” a dark albedo with more samples.
  • Save the high-sample job for the approved frame only.

Organized libraries make this checklist faster: organize your texture library. More production notes on the blog.

Frequently asked questions

What reduces render time the most in PBR scenes?

Lower preview resolution, fewer lights, smaller distant textures, and a denoiser so you can cut samples.

Is denoising a replacement for samples?

It lets you use fewer samples, but too few smears metal edges. Find the lowest count that still holds highlights.

Do 8K PBR textures always look better?

No. Match map size to how many pixels the surface covers. Extra resolution often only costs VRAM.