Watse D'haene’s Post

AI video is moving into the feed. The edit still isn’t. X is testing text-to-video in the post composer and Grok shipped Extend and Upscale. Making a clip is becoming a button — making a sequence is not. For a couple of years the hard part of AI video was getting a usable clip at all. That part is quietly ending. Reports this month have X testing text-to-video straight from the post composer, Grok animating a still from your phone, and the iOS app adding two pointed editing features — Extend and Upscale. Generation is turning into a button in the feed, not a workflow you set aside an evening for. That is genuinely good, and it is also the exact moment the bottleneck moves. When everyone can produce four seconds on demand, four seconds stops being the deliverable. The sequence does. The undervalued skill — putting shots in an order that holds together — is the one about to matter most. You can see it in the features themselves. “Extend” exists because a clip ends before the moment does. “Upscale” exists because the first pass is rarely the final pass. Both are admissions that the raw generation is an ingredient, not the dish. The work is still downstream, in the cut. We have written about the parts that survive this shift: the habits that separate keepers from near-misses, and how to rank models by whether they hand you a shot you can actually use rather than a pretty demo. None of that changes because the generate button got closer. If anything it matters more, because the volume is about to go up. So when the feed hands you a pile of clips, the question is the same as it has always been in editing: what order, what trims, where the cut lands. That part does not happen in a composer. It happens on a timeline. https://lnkd.in/eUwRCgTD

Excuse me, but are you high? "Overall I was surprised by how many of these 120 games use generative AI for music and assets. I had assumed that the vast majority would be for translation and store page images. The former raises quality control questions and is obviously a disaster for actual humans who are employed to translate games, but it's arguably less egregious compared to machine-made music, textures and narrative." Link to the article in comments because… algorithm. #gameloc #hottakes #areyouhigh

A cow with wings. A bear that has you round for tea. A hare as tall as you, standing in the lane. To everyone else they're a kid's tall tales. To him they just happened. I made a one minute short around that idea. A boy on the train home tells his mum about his summer at his grandparents', where the ordinary stuff and the impossible stuff carry exactly the same weight. I shot it to look like a faded Super 8 home movie, all grain and gate weave. Under the story it was really a test of Seedance 2.0 Mini, ByteDance model, run inside Dreamina AI. What I actually came away with was a workflow, not a render. A few honest notes. Mini is fast and cheap but it doesn't always hold up. The fiddly things like smoke, fire and subtle facial emotion tend to fall apart on it. So I stopped fighting it on the expensive model. Now I block every shot on Seedance 2.0 Mini or Seedance Fast 480p to lock the camera move, the timing and the framing for almost nothing, then I only send the keepers through full Seedance 2.0 for the final quality. The premium model never touches anything that isn't already working. Why it's worth caring about. On one short test it's a few euros, easy to ignore. But run it like a studio, say twenty films a month at realistic take counts, and that one habit saves somewhere around 16,000 euros a year. It also drops premium renders from roughly 4,800 down to about 600. Same films, half the spend. The stack was Midjourney V8 for the stills, Dreamina AI for the Seedance renders, and DaVinci Resolve for the edit. Block the shot for cents. Spend the credits only where they show up on screen. Turns out the optimisation wasn't in the model, it was in the order you use them. #Seedance #AIFilmmaking #GenerativeAI #ProductionWorkflow #AIVideo

I built this video ad with AI from scratch — no camera, no studio. Here's the pipeline: → GPT-Image-2 for storyboard frames → Seedance 2 to animate them into video → ElevenLabs for the voice performance → Lip-sync correction pass to tighten it up Real client brief. Real deliverable. This is what AI-first creative production looks like in 2026.

I built this video ad with AI from scratch — no camera, no studio. Here's the pipeline: → GPT-Image-2 for storyboard frames → Seedance 2 to animate them into video → ElevenLabs for the voice performance → Lip-sync correction pass to tighten it up Real client brief. Real deliverable. This is what AI-first creative production looks like in 2026.

🌐 Horizon Mapping Series [013] 📄 An LLM Agent Framework for Automated 3D Cutscene Generation [Unreal 5] ⚡ TLDR: The Cutscene Agent framework uses large language models to automate the generation of 3D cutscenes in video games and interactive media by integrating with game engines and orchestrating multiple specialist agents. ⚙️ Key technical contribution is the development of a bidirectional integration between LLM agents and game engines, enabling closed-loop generation of editable engine-native cinematic assets through continuous observation of real-time scene state. 🔍 The research has relevance to game engines as it uses them for cutscene generation, and to generative AI as it utilizes large language models for automation, while also having potential implications for XR as interactive media is a key application area. Paper, code & demo: https://lnkd.in/dE-5nrRQ

Love Kohji Nagata but respectfully disagree on this one. Sure, you can keep building on UE5. But all the new features will be on UE6. And UE6 will be agentic. But so will Unreal alternatives. So if you are making a game in 2027 like Expedition 33 that doesn't rely on Unreal's higher end aspects, you have to decide on whether you want to bank on something like Blueprints which you have been told is going away. Or on another technology that's not as boutique as Verse that has a clear runway ahead of it.

Epic is reworking Unreal around Verse for UE6, and the panel split. Chris: "It's not the Blueprints thing that bothers me, it's that they're replacing the scene graph with a Verse-based one. You're standing in front of your developers saying this thing you think sucks is now the future." Kohji: "From the guy who said fire anyone who won't use AI: Unreal is doing the same thing, simplifying it for people who don't want to code, and now it's an outrage? Nothing's stopping you from using the current engine." Full episode: https://lnkd.in/gb7dprFg

What actually works in AI video right now After enough generations you stop chasing the perfect prompt and start noticing what the keepers have in common. None of it is a secret setting. It is mostly discipline about scope. Keep shots short and single-purpose One action per clip. A character sits down. A camera pushes in. The longer you ask a model to sustain a moment, the more time it has to lose the plot. Two good seconds beat eight that fall apart, and a sequence is built from short shots anyway. Decide the frame before you generate The generations that work tend to start from a clear idea of the composition — where the subject sits, where the camera is, which way it moves. Blocking that out first, even roughly, gives the model a target instead of a vibe. This is the whole point of doing previs before you spend a token on a render. Lock a look and repeat it Consistency reads as intent. Pick a lens feel, a palette, and a lighting direction, and carry the same language across every prompt in a scene. Clips that share a look cut together even when they were generated hours apart. Generate more than you need, then be ruthless Plan to throw most of it away. The keeper rate on directed shots is real but it is not 100%, so generate in batches and cut hard. The skill is not making every clip perfect; it is recognizing the three that are and ignoring the rest. Cut early, cut often The sequence tells you what you are missing — a reverse angle, a beat of breathing room, a tighter insert. You only see those gaps once the clips are on a timeline next to each other. The edit is not the last step; it is how you find out what to generate next. That last habit is the one most people skip. Put your clips in order before you decide the project is not working — half the time the problem was sequence, not footage. Original post: https://lnkd.in/ew4KCMZi

Mocha Pro 2026.5 Adds Automatic Point Tracking, Advanced AI Matte Refinement, More: Boris FX has released Mocha Pro 2026.5. This version includes new AI-driven tracking and masking workflows built for VFX artists and compositors. Key features include Point Track ML and improved... The post Mocha Pro 2026.5 Adds Automatic Point Tracking, Advanced AI Matte Refinement, More appeared first on . https://lnkd.in/gD-nE3nv More: Toolfarm.com