Camera lens profiles

A perfect render does not look like camera footage, and a model trained only on perfect renders notices. Lens profiles apply a measured amount of the wrong to your images: haze, grain, dirt, glare, compression. This page lists every profile that ships.

Applied after capture, which changes everything#

Lens simulation runs offline, in the report and dataset exporters, not in Unreal. Three consequences worth knowing:

Changing the profile does not require re-rendering. Re-run nameframe dataset against the same dump.
You can export the same capture several times with different profiles and compare what the model does with each.
Labels are unaffected. They are derived from the ID buffer, which the lens never touches.

# in a recipe
sensor:
  enabled: true
  profile: realistic_drone
  strength: 1.0
  save_clean_images: true

# or per export
nameframe dataset _local/dumps/beach _local/datasets/beach --task box --lens-profile realistic_drone --lens-intensity 1.0 --include-clean-images

Keep the clean copy

save_clean_images, or --include-clean-images, writes the untouched RGB next to the processed images. It costs disk and saves you from re-rendering when you want to see what the lens actually did.

Intensity#

0 leaves the image clean, 1.0 is the profile as designed, and up to 2 pushes it hard. Values above 1 are useful for stress-testing a model, and less useful for training data that is meant to look plausible.

The catalogue#

27 distinct looks, reachable under 40 names plus 4 random aliases. An alias is an exact synonym: it resolves to the same parameters.

Profile	Also accepted as	What it looks like
`none`	`clean`	No treatment. The render passes through untouched.
`night_vision`	`infrared_green`	Green image intensifier: heavy gain, noise and bloom.
`thermal_white_hot`	`thermal`	Thermal imager, white-hot palette.
`sepia_vintage`	`vintage`	Aged film stock: warm cast, soft edges, heavy grain.
`golden_hour`	none	Low warm sun, long glare, mild haze.
`harsh_noon`	`blown_highlights`	Blown highlights and hard contrast, as at midday.
`night_street`	`sodium_night`	Sodium street lighting, orange cast, low-light noise.
`cold_cine`	`teal_cine`	Teal cinematic grade, cool and slightly desaturated.
`action_cam`	`gopro`	Wide sports camera: strong contrast, chromatic fringing.
`vhs_analog`	`vhs`	Analogue tape: bleed, scanline noise, heavy compression.
`cheap_webcam`	`webcam`	Low-cost sensor: soft, noisy, aggressively compressed.
`underexposed_noise`	`low_light_noise`	Shot below the sensor's usable range.
`realistic_drone`	`drone_camera`, `runpod_realistic`	The default aerial look: mild haze, light dirt, moderate grain.
`atmospheric`	none	Distance haze with a cool cast.
`dusty_aerial`	none	Aerial through dust: strong haze, warm cast, dirty optics.
`cool_haze`	none	Cold desaturated haze, low contrast.
`sharp_surveillance`	none	Clean fixed camera: sharp, low grain, high quality.
`cctv`	none	Older fixed camera: desaturated, blocky, low bitrate.
`dashcam`	none	Vehicle camera: wide, contrasty, visible compression.
`bodycam`	none	Worn camera: soft, noisy, motion-tolerant.
`phone_camera`	none	Modern phone: heavy sharpening and saturation.
`old_security_camera`	none	Degraded fixed camera: washed out and blocky.
`dirty_industrial_lens`	none	Grease, dust and scratches on the front element.
`rainy_lens`	none	Water on the lens: streaks, smearing, glare.
`low_sun_dirty`	none	Low sun straight into a dirty lens: heavy glare and ghosting.
`dirty_lens_low_contrast`	none	Neglected optics: low contrast, dirt, dust, scratches.
`fogged_camera`	none	Condensation on the element: soft, hazy, desaturated.

Random and weighted selection#

These four names all mean the same thing, pick one per frame: random, weighted, sensor_random, camera_random.

Without explicit weights, this is the built-in distribution. Weights are relative, so they do not have to add up to anything in particular.

Profile	Default weight
`dusty_aerial`	20
`dirty_lens_low_contrast`	18
`realistic_drone`	16
`atmospheric`	12
`golden_hour`	11
`dashcam`	10
`rainy_lens`	10
`cold_cine`	9
`harsh_noon`	9
`action_cam`	8
`dirty_industrial_lens`	8
`low_sun_dirty`	8
`night_street`	7
`fogged_camera`	7
`cctv`	6
`night_vision`	6
`sepia_vintage`	5
`bodycam`	5
`vhs_analog`	5
`underexposed_noise`	5
`cheap_webcam`	4
`old_security_camera`	4
`thermal_white_hot`	4
`phone_camera`	4

What a profile is made of#

Every profile is a set of these values. You can override individual ones through sensor.effects in a recipe, for when a stock profile is close but not exactly what you want.

Parameter	Effect
`chromatic`	Chromatic aberration, colour fringing towards the edges
`bloom_threshold`, `bloom_intensity`, `bloom_radius`	Highlight bloom
`vignette`, `vignette_power`	Corner darkening and its falloff
`grain`, `grain_chroma`	Luminance and colour sensor noise
`contrast`, `exposure`	Tone curve and overall brightness
`warmth`, `tint`	Colour temperature and per-channel tint
`softness`	Optical blur, a lens that is not quite sharp
`desaturate`	Colour strength reduction
`haze`, `haze_gradient`, `haze_color`	Atmospheric veil and its vertical gradient
`dirt`, `dirt_blobs`, `dust`, `scratches`, `grease`	Contamination on the front element
`glare`, `glare_ghosts`	Veiling glare and lens-flare ghosts
`blockiness`, `jpeg_quality`	Compression artefacts

Choosing one#

Match the profile to the camera you are going to deploy on, not to what looks good.

Deploying on	Start with
Drone or aerial imagery	`realistic_drone`, or `dusty_aerial` in dry climates
Fixed security camera	`cctv`, or `sharp_surveillance` for a modern one
Vehicle-mounted	`dashcam`, plus `rainy_lens` in the mix
Worn or handheld	`bodycam` or `phone_camera`
Night operations	`night_street`, `night_vision`, `underexposed_noise`
Thermal sensors	`thermal_white_hot`
Mixed or unknown	`random` with weights matching your fleet

Weighted mixes

Real fleets are not one camera. Weight the mix to match yours:

sensor:
  profile: random
  profile_weights: "realistic_drone:40,dusty_aerial:25,rainy_lens:20,fogged_camera:15"

Or on the CLI:

nameframe dataset _local/dumps/x _local/datasets/x --lens-profile random --lens-profile-weights "realistic_drone:40,dashcam:25,rainy_lens:20,cctv:15"

Tuning one knob

When a stock profile is close but not right, override individual parameters instead of building a new one:

sensor:
  profile: cctv
  effects:
    grain: 8.0
    jpeg_quality: 55

Do not over-apply

Lens simulation is the last thing you should add and the easiest to overdo. Heavy dirt and grain on top of an already-hazy scene produces images no real camera would output, and the model learns the artefact rather than the object. Add it after the camera and weather variation already work, and look at the result.