••• | Black Forest Labs, the group behind the FLUX image models, released a video upscaler on 20 August. You hand it a small, rough clip and it hands back a larger, sharper one, up to 4K. It runs through their own service and through the playground on their site, and ComfyUI added a ready made workflow for it the same day. The feature is ordinary. What it changes about the work is not. Video has always been priced so that the expensive part happens first: you pay for a big render, look at it, and throw most of them away. Now the order can flip. Make ten small takes at low cost, watch them, and pay for size only on the one you keep. That is the whole idea, and there is one setting that will ruin it if nobody tells you about it. | ONE · THE SETTING It is on by default, and it changes faces. | There is a control called creativity, and it has two useful positions. At zero, the tool, in their words, preserves the source exactly. Their own guidance says to use that when identity matters: faces, products, brand assets, footage of real people. At one, it restores and invents fine detail more aggressively. Their documentation is honest about the cost of that, in the same sentence: faces and products can drift. Not blur. Drift. The person comes back looking slightly like someone else. The setting defaults to one. So the mode that quietly redraws a face is the mode you get if you change nothing, and you will only notice on the clip you already decided to keep. If there is a human being or a client's product in the frame, set creativity to zero yourself, every time. Nothing in the interface will remind you. | TWO · THE LIMITS Twenty seconds, and you pay by the pixel. | Four numbers before you plan around this. A clip can be at most twenty seconds and 50 megabytes. The largest thing you can put in is 2560 by 1440, so it upscales toward 4K and will not accept 4K going in. And it multiplies by between 1.5 and 3, with 2 as the default. The charge is per megapixel of every second, which is an unfamiliar shape until you see it in money. Their own worked example: a ten second clip at 1080p costs $1.39 in the exact mode, or $1.98 in the inventive one. Push the same ten seconds to 4K and you are into single figures. A clip that gets rejected is not charged, which matters more than it sounds when you are feeding it output from another tool. | THREE · ALSO THIS WEEK Eight camera moves you can name. | On 19 August, Lightricks added camera direction to the LTX mode that builds video out of an audio track. Until now you fed it a song and accepted whatever shot you were given. It is listed in their change log alongside frame rate and a setting that pins the final frame. The move names are a fixed list rather than open description, so they are worth writing down: dolly in, dolly out, dolly left, dolly right, jib up, jib down, static, and focus shift. Jib up and jib down are the crane moves, lifting or dropping the whole camera rather than tilting it. Focus shift moves the sharp plane between subjects without moving the camera at all. Eight is a small vocabulary, and a small vocabulary you can actually rely on is worth more than an open text box that ignores you. | PROMPT OF THE DAY The one that stops it inventing. | The upscaler takes an optional line of text that steers what detail it adds. Most people leave it empty, which hands the decision back to the model. Use this when there is a person or a product in the shot, together with creativity set to zero. Sharpen detail that is already present. Keep skin texture, fabric weave, and any printed text exactly as they appear in the source. Do not add freckles, stitching, jewellery, or logo detail that is not already visible. Do not smooth skin. Do not change the shape of any face. The four sentences beginning with Do not are the ones doing the work. A steering line made only of positive instructions reads to the model as permission to improve things, and improving a face is exactly the failure you are trying to avoid. Their own advice points the same way and is worth repeating: start from the least compressed source you have. An upscaler will happily sharpen compression damage into something that looks deliberate. | THE NEXT WORKING NOTE Stop rendering big while you are still deciding. | The habit this suggests costs nothing and does not depend on this one tool. When you are still choosing between ideas, generate at the smallest size that lets you judge the thing you are judging. Framing, timing and motion are all legible at low resolution. Detail is the last question, not the first, and it is now the only part you have to pay a premium for. Keep the small versions too. When a client asks for a different edit in two months, the small takes you already rejected are a library, and the one you finished is only one answer out of ten. | ••• I am putting together an upscale kit: the settings to change before you run a clip, six steering lines for different subjects, and a short note on which kinds of source footage come back worse than they went in. Want it when it is ready? Reply with send me the upscale kit and I will get it to you. | A QUESTION FOR YOU What are you still rendering at full size too early? I did it for a year with images before it occurred to me that I was paying to look at things I was going to delete. I suspect most of us have one step like that left. If this was useful, forward it to someone who works in video. | Until next time, Luxe Prompting | Luxe Prompting AI Systems and Prompt Craft for Creators |
|
