“I used GPT-4 to describe itself. Then I used its description to generate an image, a video based on this image and a soundtrack.
Tools I used: GPT-4, Midjourney, Kaiber AI, Mubert, RunwayML
This is the description I used that GPT-4 had of itself as a prompt to text-to-image, image-to-video, and text-to-music. I put the video and sound together in RunwayML.
GPT-4 described itself as: “Imagine a sleek, metallic sphere with a smooth surface, representing the vast knowledge contained within the model. The sphere emits a soft, pulsating glow that shifts between various colors, symbolizing the dynamic nature of the AI as it processes information and generates responses. The sphere appears to float in a digital environment, surrounded by streams of data and code, reflecting the complex algorithms and computing power behind the AI”
This page compares images rendered in Arnold using spectral rendering and different sets of colourspace primaries: Rec.709, Rec.2020, ACES and DCI-P3. The SPD data for the GretagMacbeth Color Checker are the measurements of Noburu Ohta, taken from Mansencal, Mauderer and Parsons (2014) colour-science.org.
“Unless you have all the relevant spectral measurements, a colour rendition chart should not be used to perform colour-correction of camera imagery but only for white balancing and relative exposure adjustments.”
“Using a colour rendition chart for colour-correction might dramatically increase error if the scene light source spectrum is different from the illuminant used to compute the colour rendition chart’s reference values.”
“other factors make using a colour rendition chart unsuitable for camera calibration:
– Uncontrolled geometry of the colour rendition chart with the incident illumination and the camera.
– Unknown sample reflectances and ageing as the colour of the samples vary with time.
– Low samples count.
– Camera noise and flare.
– Etc…
“Those issues are well understood in the VFX industry, and when receiving plates, we almost exclusively use colour rendition charts to white balance and perform relative exposure adjustments, i.e. plate neutralisation.”
Spectralon is a teflon-based pressed powderthat comes closest to being a pure Lambertian diffuse material that reflects 100% of all light. If we take an HDR photograph of the Spectralon alongside the material to be measured, we can derive thediffuse albedo of that material.
The process to capture diffuse reflectance is very similar to the one outlined by Hable.
1. We put a linear polarizing filter in front of the camera lens and a second linear polarizing filterin front of a modeling light or a flash such that the two filters are oriented perpendicular to eachother, i.e. cross polarized.
2. We place Spectralon close to and parallel with the material we are capturing and take brack-eted shots of the setup7. Typically, we’ll take nine photographs, from -4EV to +4EV in 1EVincrements.
3. We convert the bracketed shots to a linear HDR image. We found that many HDR packagesdo not produce an HDR image in which the pixel values are linear. PTGui is an example of apackage which does generate a linear HDR image. At this point, because of the cross polarization,the image is one of surface diffuse response.
4. We open the file in Photoshop and normalize the image by color picking the Spectralon, filling anew layer with that color and setting that layer to “Divide”. This sets the Spectralon to 1 in theimage. All other color values are relative to this so we can consider them as diffuse albedo.
One problem with sRGB is that in a gradient between blue and white, it becomes a bit purple in the middle of the transition. That’s because sRGB really isn’t created to mimic how the eye sees colors; rather, it is based on how CRT monitors work. That means it works with certain frequencies of red, green, and blue, and also the non-linear coding called gamma. It’s a miracle it works as well as it does, but it’s not connected to color perception. When using those tools, you sometimes get surprising results, like purple in the gradient.
There were also attempts to create simple models matching human perception based on XYZ, but as it turned out, it’s not possible to model all color vision that way. Perception of color is incredibly complex and depends, among other things, on whether it is dark or light in the room and the background color it is against. When you look at a photograph, it also depends on what you think the color of the light source is. The dress is a typical example of color vision being very context-dependent. It is almost impossible to model this perfectly.
I based Oklab on two other color spaces, CIECAM16 and IPT. I used the lightness and saturation prediction from CIECAM16, which is a color appearance model, as a target. I actually wanted to use the datasets used to create CIECAM16, but I couldn’t find them.
IPT was designed to have better hue uniformity. In experiments, they asked people to match light and dark colors, saturated and unsaturated colors, which resulted in a dataset for which colors, subjectively, have the same hue. IPT has a few other issues but is the basis for hue in Oklab.
In the Munsell color system, colors are described with three parameters, designed to match the perceived appearance of colors: Hue, Chroma and Value. The parameters are designed to be independent and each have a uniform scale. This results in a color solid with an irregular shape. The parameters are designed to be independent and each have a uniform scale. This results in a color solid with an irregular shape. Modern color spaces and models, such as CIELAB, Cam16 and Björn Ottosson own Oklab, are very similar in their construction.
By far the most used color spaces today for color picking are HSL and HSV, two representations introduced in the classic 1978 paper “Color Spaces for Computer Graphics”. HSL and HSV designed to roughly correlate with perceptual color properties while being very simple and cheap to compute.
Today HSL and HSV are most commonly used together with the sRGB color space.
One of the main advantages of HSL and HSV over the different Lab color spaces is that they map the sRGB gamut to a cylinder. This makes them easy to use since all parameters can be changed independently, without the risk of creating colors outside of the target gamut.
The main drawback on the other hand is that their properties don’t match human perception particularly well.
Reconciling these conflicting goals perfectly isn’t possible, but given that HSV and HSL don’t use anything derived from experiments relating to human perception, creating something that makes a better tradeoff does not seem unreasonable.
With this new lightness estimate, we are ready to look into the construction of Okhsv and Okhsl.
import math,sys
def Exposure2Intensity(exposure):
exp = float(exposure)
result = math.pow(2,exp)
print(result)
Exposure2Intensity(0)
def Intensity2Exposure(intensity):
inarg = float(intensity)
if inarg == 0:
print("Exposure of zero intensity is undefined.")
return
if inarg < 1e-323:
inarg = max(inarg, 1e-323)
print("Exposure of negative intensities is undefined. Clamping to a very small value instead (1e-323)")
result = math.log(inarg, 2)
print(result)
Intensity2Exposure(0.1)
Why Exposure?
Exposure is a stop value that multiplies the intensity by 2 to the power of the stop. Increasing exposure by 1 results in double the amount of light.
Artists think in “stops.” Doubling or halving brightness is easy math and common in grading and look-dev. Exposure counts doublings in whole stops:
+1 stop = ×2 brightness
−1 stop = ×0.5 brightness
This gives perceptually even controls across both bright and dark values.
Why Intensity?
Intensity is linear. It’s what render engines and compositors expect when:
Summing values
Averaging pixels
Multiplying or filtering pixel data
Use intensity when you need the actual math on pixel/light data.
Formulas (from your Python)
Intensity from exposure: intensity = 2**exposure
Exposure from intensity: exposure = log₂(intensity)
Guardrails:
Intensity must be > 0 to compute exposure.
If intensity = 0 → exposure is undefined.
Clamp tiny values (e.g. 1e−323) before using log₂.
Use Exposure (stops) when…
You want artist-friendly sliders (−5…+5 stops)
Adjusting look-dev or grading in even stops
Matching plates with quick ±1 stop tweaks
Tweening brightness changes smoothly across ranges
Use Intensity (linear) when…
Storing raw pixel/light values
Multiplying textures or lights by a gain
Performing sums, averages, and filters
Feeding values to render engines expecting linear data
Examples
+2 stops → 2**2 = 4.0 (×4)
+1 stop → 2**1 = 2.0 (×2)
0 stop → 2**0 = 1.0 (×1)
−1 stop → 2**(−1) = 0.5 (×0.5)
−2 stops → 2**(−2) = 0.25 (×0.25)
Intensity 0.1 → exposure = log₂(0.1) ≈ −3.32
Rule of thumb
Think in stops (exposure) for controls and matching. Compute in linear (intensity) for rendering and math.
An exposure stop is a unit measurement of Exposure as such it provides a universal linear scale to measure the increase and decrease in light, exposed to the image sensor, due to changes in shutter speed, iso and f-stop.
+-1 stop is a doubling or halving of the amount of light let in when taking a photo
1 EV (exposure value) is just another way to say one stop of exposure change.
Same applies to shutter speed, iso and aperture.
Doubling or halving your shutter speed produces an increase or decrease of 1 stop of exposure.
Doubling or halving your iso speed produces an increase or decrease of 1 stop of exposure.
Divesh Naidoo: The video below was made with a live in-camera preview and auto-exposure matching, no camera solve, no HDRI capture and no manual compositing setup. Using the new Simulon phone app.
DISCLAIMER – Links and images on this website may be protected by the respective owners’ copyright. All data submitted by users through this site shall be treated as freely available to share.
