§ · Tools
Choose a tool by the job it has to do. These dispatches include practical workflows, release analysis and opinion. An analysis of a new feature is not a production test; check the scope in each article. 40 recent dispatches are listed below.
A camera instruction and a scene description solve different problems. Start with camera movement, focus and lighting vocabulary. For staging before generation, read the Blender blockout analysis.
Choose how much of the process you need to expose. The ComfyUI App Mode guide looks at working with prepared workflows. The Higgsfield Content Scoring analysis asks what a score can tell you after a shot is made.
Generating an image does not settle where the cut belongs. Read the division between automated editing tasks and take selection, then the Remotion guide to repeatable video variants. For completed work, visit the project cases.
Download REA free: the skill that lets your agent take any app apart and explain it. I tested it, and here is why the app on demand worries me.
Five Claude skills reviewed from their own files: a Word redliner, a palette tool, a ComfyUI kit, a NotebookLM driver and a screenwriting pack, with limits.
NVIDIA found a flaw in one in four agent skills it scanned, then listed what its scanner cannot read. Five new Claude skills and the check each leaves you.
Five Claude skills opened on September 17. One sends your photo slots to a server answering 503. What each one does in the step you never chose.
What the Toninho Sagatiba draft reviews caught, how I separate a film's creative rules from its pitch, and a practical review prompt with a free skill.
Genjutsu recasts a shot and keeps the camera move. What it costs per clip, the cheap-first order I run, and the input limits Higgsfield now agrees on.
Five Claude skills opened repo by repo on September 7. One hides a 4K video exporter in neither manual. The file check I run before installing.
The Sora 2 API is removed on September 24 and OpenAI named no replacement. What a working pipeline loses, and why the swap should take an afternoon.
Five Claude skills opened and counted on August 31, plus the number I promised and did not want: fifteen of thirty five patterns were in my own posts.
Google's new video model charges ten times more for 4K than for a 360p draft. Its own API docs call that 4K an upscale. Here is how I read the tier list.
Junkyard King Chapter 2 held six characters across fifteen minutes. The method behind it was built before a single frame was ever generated.
AI drama is the one format where generated video is not a compromise. A step-by-step method to build a series that holds, from character sheets to paywall.
Thirty second clips with audio in one pass, and no weights to download. Why the licence decides more of a working pipeline than any benchmark does.
The DALL-E GPT is off inside ChatGPT. What dies is the prompt record, not the picture, and a six step archive rule from post keeps yours alive.
Five Claude skills opened and counted on August 22, not recalled from memory, plus the copy on my own drive that had sat five months out of date.
ByteDance signed with the MPA and the guardrails are output filters. They rarely stop the person copying Iron Man. They stop the shot due on Friday.
When the answer is a number, the chatbot keeps the click. So I built a calculator instead, with every data rate traced back to a published document.
LTX 2.5 returns several connected shots in one pass, so the model picks your cut. Where to run it free, what changed since launch, and when to cut by hand.
The EU marking duty is live and the vendors did their half. The other half breaks the first time you export a master. What changes in delivery, step by step.
Seedance 2.5 Draft Mode lets you re-roll cheap 480p takes and render only the keeper at 1080p. When native 1080p pays off, and the exact steps to run it.
How to use Higgsfield Cinema Studio 4.0 and its Emotion Wheel like a director: 30 second takes, 50 references, and the note no actor or model can play well.
Premiere's AI Assistant, Object Mask and silence removal already work. Choosing which take carries the scene is still the call no model makes for you.
Every executor gets a harsh verifier, and the loop only stops when all of them are wowed. I show the prompt structure, the four fixes and where to cap it.
A labor lawyer in Santos got a site where the camera flies into a port at night. The pipeline behind it, plus the one rule that breaks the whole effect.
Edgar Pêra mixes archive and generated imagery, then plants clues so you can tell them apart. Why that is a directing decision and not a disclaimer.
A $1,000,000 prize pool, an Academy Award winning jury, and one clause worth reading twice: when submissions close, every project file becomes public.
The 50 reference slots, the 30 second pass and region editing, read as a call sheet. Where the model pays you back and where it burns credits twice.
Grok Imagine locks up to seven references per shot. I direct commercials, and I can tell you what each lock buys and the shot it quietly kills.
OpenAI closed Sora and told Disney under an hour before the public. What that costs anyone whose pipeline runs through a single rented endpoint.
Two Bloomberg columns, three days apart, on the same week in AI cinema. One shows the raise before the film. The other shows a boy with a camcorder.
Autodesk put a 3D stage, camera tracks and blocking into Flow Studio. What changes on a shot when you stop arguing with a model and place the lens yourself.
Article 50(4) limits the duty to a disclosure that does not hamper the work, and a cleanup pass is not a generated performance. Where that line falls for you.
Fifteen seconds of 2K with stereo sound from a single generation, around $0.13 per second. What collapsing image and sound into one pass costs the craft.
Reply named its winner in Venice on September 6, 2026. What the ten finalists had that the other 2,990 entries did not, named without flattery.
Seven agents take one prompt from script to finished short: storyboard, character design, music, synthesis. The decision the pipeline hands back to you.
Qwen-Image-3.0 arrived with no weights, no benchmark and no technical report. For anyone working to a client deadline, the missing paperwork is the review.
TikTok Shop ships an AI Video Maker and SharkNinja bans it in the same week. What a working AI UGC director does when platform and brand disagree.
One hired 1,000 rotoscope artists. The other joined an AI storyboard company. Both protect the same thing, and it is not the tool they picked.
Thirteen minutes made with Seedance 2.0, faces and voices from 32 real people, then a studio deal. What the choices in the edit say about authorship.
FLUX 3 Video makes dialogue and room tone in the same pass as the picture. Is it out, how to get access, and why sound is the right test, not 20 seconds.