This week we're releasing 2 new ComfyUI templates, which integrate LTX's open source HDR model/LoRA, allowing you to add bit depth to your existing videos, or create "16-bit" videos from scratch.

For those of you who aren't familiar with bit-depth as a concept, it's basically the range of colour that a video can store. Imagine that you walk into a darkened room - it may initially appear black, but as your pupils open up to allow more light in, you can see more and more information. So your eyes and brain can see a greater range of colour, depending on the lighting conditions.

More bit depth allows a video to "open its pupils" (or close them). A black video at 8-bit can only be a black video. At higher bit depth (typically 10 or 12-bit), you can brighten it to find the information hidden in the blacks.

That's just an example. The same happens with whites, and with colours - basically you're storing a greater range of colour, giving you more room to play with the image in post production.

Working at a higher bit depth isn't essential, unless you're working professionally, in which case it's a must. If Netflix asks you to make the shot brighter so you can see more of it, you can't say "No". Well, someone said no to HBO on House of the Dragon, because that show came out dark, but I digress.

Since a higher bit depth is essential for professionals, having a model that allows you to work in a 10-bit workflow whilst using AI is a big deal - and that's exactly what this LTX model does...

OR DOES IT?

Eh.. sort of.

I'm always ready to throw a bucket of icy water on AI hype nonsense, because it's out of control, but with LTX I'm going to tuck it into bed, and make it a nice cup of hot chocolate. It's clever as hell, and in the right conditions works pretty nicely, although I'm still going to dunk it's toes into a glass of fairly cold water after, because honestly it's quite limited.

Here's the low down:

The video is actually STILL generated at 8-bit.

AI video models generate colour values within a fixed range - essentially a numbered scale from 0 to 255. That's what 8-bit means - 256 colours in red, 256 in green, and 256 in blue. That means you can mix them to achieve 256x256x256... 16.7 million different colours.

It's a lot, but it's not enough to capture the full spectrum of human vision. 10-bit allows for 1024x1024x1024... over a billion colours.

That's much better, but AI models are trained at 8-bit, and that's already insanely computationally expensive. To create a model that generates at 10-bit would be madness.

So here's what's happening in LTX's model (white paper is here if you're feeling brainy).

It's generating an image at 8-bit, but in logarithmic space. That's a term you'll hear a LOT in film, so let's define what we're talking about.

From the point of view of human vision, the difference between 0 candles and 1 candle is huge. BUT the difference between 1000 and 1001 candles is impossible to spot. Our vision is logarithmic, and the easiest way to measure the increase or decrease in light is in stops, which is the doubling or halving of light. One stop brighter - double the amount of light, one stop darker, half.

But from a computer's point of view which works in "linear space", the difference between 0 and 1 or 1000 and 1001 candles is the same - it's 1 candle.

If you store a digital image in linear space (very common), all colour is treated equally. If you store it in logarithmic space, you give more room to the areas that matter more to humans.

If you store a logarithmic image in 8-bit, you can still only record the same amount of colours (16.7 million), BUT you're dedicating more space to the colours that matter (to a human). So from the human's point of view, you're storing more colour, even though technically you're not.

Now what happens if you want that colour in a linear space? Imagine the curve was a piece of string - to make it straight, you need to stretch it out, right? Increasing the range.

So while a video in this model is still generated at 8-bit, it's in logarithmic space, trained on logarithmic images (actually HDRI images from Poly Haven and the open-source short film Tears of Steel, which gave the model real-world human motion and natural lighting to learn from) - it is then decoded and transformed (stretched out to be straight), meaning it's now technically longer, and therefore needs more bit depth to store that extra information.

I talk a lot about the price of going into latent space (the world where AI imagery is pulled from). Since it's a world without pixels or polygons, your image will always change when you come in and out of it. Kind of like translating backwards and forwards between two different languages - the general meaning is the same, but the more you do it, the more you lose.

For this LTX model, you'll pay a reasonable cost when you step into latent space. The image you get out should have the same proportions and composition as the one that went in, but the colours will have shifted.

Where you might find it more useful is if you're filming yourself and you don't have the equipment or experience to nail the exposure or to film at a high bit depth (most phones and tablets won't, unless they're high end). That's a situation where LTX could get you out of a bind!