Samples and bounces, settled
A clean interior still in Blender Cycles usually needs 64–128 samples with denoising and 4–8 diffuse bounces — not the 4096-sample default. This reference lays out the numbers that decide whether a render looks photographic or plastic, what they cost in render time, and how to read the failures when something looks wrong.
| Setting | Working value |
|---|---|
| Final samples, denoised still | 64–128 samples |
| Denoising floor | ≈32 samples |
| Diffuse bounces | 4–8 |
| Glossy bounces | 2–4 |
| Exposure, daylit interior | −0.7 to −0.3 EV |
| Exterior plate strength | 0.2–0.5 |
| Focal length, full-frame equivalent | 8–24 mm |
| Delivery resolution | 2000–4000 px long edge |
64–128 samples with denoising, 4–8 diffuse bounces and 2–4 glossy bounces — that is the whole preset for most interior stills.
Cycles ships with 4096 samples for final renders and 1024 for viewport previews. Those defaults are written for worst-case scenes — caustics, heavy volumetrics, deep indirect light — and treating them as a target is how a still that should take minutes turns into an overnight job.
With the OpenImageDenoise denoiser and its Albedo and Normal passes enabled, 64–128 samples is usually enough for an interior still. The denoiser uses those auxiliary passes to tell surface detail apart from shading noise, which is why the passes matter as much as the sample count itself.
Bounces are the second half of the preset. Typical interior counts are 4–8 diffuse and 2–4 glossy; the diffuse value carries indirect light into corners, and the glossy value keeps reflections alive on metal and glass.
The view-transform change matters because exposure habits built on Filmic do not transfer one-to-one to AgX.
The shipping default is a ceiling for pathological scenes, and render time scales almost linearly with sample count.
A 1920×1080 interior still at 512 samples with denoising takes from a few minutes to a few tens of minutes on a current GPU. The same scene at 4096 samples without denoising runs 20–60 times longer — the difference between an afternoon of iterations and a queue you check in the morning.
Because time scales with samples, every doubling costs a doubling of minutes but buys less visible improvement once the denoiser has enough signal to work with. The skill is finding the floor where the denoiser still behaves, not climbing toward the default.
Viewport previews default to 1024 samples, which is why the interactive view can look cleaner than a low-sample final render with denoising switched off — the two numbers answer different questions.
Below about 32 samples, OpenImageDenoise stops removing noise and starts inventing texture.
The denoiser needs enough traced paths to separate signal from grain. With the Albedo and Normal passes enabled it can distinguish shading noise from genuine surface detail, which is exactly what makes the 64–128 sample range viable for an interior still.
Below roughly 32 samples that separation fails. Fine texture — hair, foliage, fabric weave — is smeared into watercolour blotches, and no extra denoiser strength brings it back, because the information was never traced in the first place.
The practical reading: if a render looks painted rather than photographic, the fix is samples, not a different denoiser. Raise the count into the 64–128 range and re-render the failing crop before committing to a full frame.
| Setting | Working value | What it controls |
|---|---|---|
| Final samples, denoised still | 64–128 samples | Noise level after OpenImageDenoise |
| Denoising floor | ≈32 samples | Below this, hair and foliage textures smear |
| Diffuse bounces | 4–8 | Indirect light reach; 1–2 causes black corners |
| Glossy bounces | 2–4 | Reflection depth on shiny materials |
| Exposure, daylit interior | −0.7 to −0.3 EV | Brightness in stops after white balance |
| Exterior plate strength | 0.2–0.5 | Window softness in an HDRI-interior hybrid |
| Focal length, full-frame equivalent | 8–24 mm | Perspective; keeps verticals from converging |
| Delivery resolution | 2000–4000 px long edge | Web banner versus full-screen hero image |
The reference draws on Blender's release notes and manual, the OpenImageDenoise documentation and the IES LM-63 photometric file standard.