32-bit color depth per channel — retains overexposed data and values beyond 0–1. Critical for HDR compositing and grading without banding artifacts.
You're working in color grading or VFX compositing and quickly realize: 8-bit is a liar. With values between 0 and 255, you have no chance of preserving subtle gradations when you triple exposure-correct a clip or stack multiple effects. This is where floating point comes in — a numerical system that stores colors not as discrete integers, but as 32-bit floating-point numbers per channel. In practice, this means you can work with values well beyond 1.0, and the software won't lose any information.
The crucial advantage lies in handling overexposure. When you're building a look in the grading suite and boost the red channel by 1.8x, the values temporarily land at 1.8 instead of 1.0 — without clipping. This isn't visible in the final result, but it allows the system to process the tonal curve more flexibly. You retain the information about what the original overexposure was and can still make corrective adjustments later. With 8-bit, that would be long gone. This flexibility is why modern HDR compositing is impossible without floating point — the extended luminance ranges of HDR require this mathematical precision.
In practice, this means: you render your VFX elements, your color-corrected masters, and your intermediate composites all in 32-bit floating-point spaces. OpenEXR is the standard for this — the format was developed precisely for this purpose. In Nuke, in DaVinci Resolve, in any professional compositing tool, you work with it internally. The visual difference compared to 16-bit integer (found in TIFF or some ProRes variants) is marginal, but the scope for corrections is exponentially larger. You notice it most when you need to perform aggressive color corrections or when multiple sequential effects must not show posterization — for example, with subtle sky gradients or hair against backlight.
A word of caution: memory intensive. A 4K sequence in 32-bit float significantly burdens RAM and hard drive space — three times that of 8-bit. Therefore, you'll eventually convert to 10-bit or 8-bit for delivery, but within the workflow itself, in all intermediate steps, you need this buffer. It's not a luxury, it's craftsmanship.