On the desk
From the desk
Roundups, op-eds, and field notes on AI video — written from actual production experience, not press releases. This is the free-form editorial and news desk.
-
ByteDance's World Model Re-Shoots a Real Movie, Krea 3 Is an Editor First, and Sora Might Come Back
ByteDance showed an unnamed world model that re-shot a Stephen Chow scene from one drawn camera path, on phone footage from a Douyin gala. Also this week: OmniHuman 2.0 (API only, not in the US), Krea 3's beta as a fast image editor that will actually edit Flamethrower Girl, The Information's report that Sora may return through a Katzenberg and Peebles startup, Pico Space Pro, and World Gacha, a free city generator that exports to Blender.
-
Gemini 4 Argon, Tested: A Week as My Daily Driver
Gemini 4 Argon is out, and I ran it as my daily driver for about a week: a Godot game, a one-shot synth, Skyrim modded overnight, a DaVinci cut of a 90-minute interview, and a way back into Planet Hell. Verdict: very good, very tenacious, not a one-shot miracle worker. The $20 Pro plan gets it, and there's no $500 tier.
-
Prism and Kandinsky 6.0, Doom Inside Skyrim, and Nano Banana 2.1
Two new open-source video models showed up in one news cycle: Prism, a 2K-with-audio preview that wants an 80 GB card and has a disputed demo reel, and Kandinsky 6.0, which runs on a 4090. Then AI-built game mods, including my own Doom-inside-Skyrim experiment, a 0.1 Nano Banana update, and a new AI assistant called Hark.
-
Kling 4 Early Look, Gemini 4 Argon, Ideogram 4.5, and Eleven v4
Kling 4 is in early access — Flash model only, 720p, Ultra annual subscribers first — and I ran it through the usual tests: lip-sync drift looks fixed, performances run a little flat, and it's maybe not the Seedance killer some had hoped for. Plus HeyGen's penny-a-second video model, Google's Gemini 4 Argon (announced, not released), Ideogram 4.5's precise editing, and Eleven v4.
-
OpenAI DevDay: dots, GPT-6.1 Sol, the Pro Pricing Shuffle, and a Few Days With Muse
OpenAI's DevDay was all about dots, ChatGPT's new always-on agents — a catch-up flex on Meta's Muse — plus GPT-6.1 Sol, an Ultrafast tier, and a Pro pricing reshuffle that won't please everyone. My dot hasn't shown up yet, so the hands-on half is a few days with Muse, where the trust question gets awkward.
-
Opus 5.5 vs GPT-6 Sol, Meta's Muse Charm, and a PixVerse R2 World You Can Talk To
Opus 5.5 and GPT-6 Sol dropped on the same day, Meta turned Muse into hardware with the Muse Charm and a $1,299 Meta VR Glasses, and PixVerse R2 lets you prompt inside a live world. On the creative side: Alibaba's unnamed November video model (the "Seedance killer" label isn't Alibaba's), a reported Kling update, Seed Audio 1.5, Fish Audio's Drama 3 preview, and Jordan Rudess's jam_bot.
-
Dreamina Platform Update: Canvas Agent, Elements, and Seedance 2.5 Preview
Not Seedance 3.0 — this is a Dreamina platform update. The canvas and its AI agent got a real overhaul, Elements finally lets characters (and voices) follow you into any prompt, and Seedance 2.5 Preview, still in preview, drafts at 480p and then re-renders the same shot at 1080. I ran three tests, including a Flamethrower Girl stress test, and found where I'd still take manual control.
-
How The Bridge Got Remade in a Day and a Half: Astra, Codex, and Runway's MCP
The making-of for The Bridge (2026): GPT-6 Astra running Codex, Runway's MCP driving Seedance 2.5, a plain-HTML production dashboard, directing by voice, and a DaVinci Resolve grade, title cards, and audio mix I never touched a mouse for. Plus the two places the agent fell down, and what the whole thing cost in usage.
-
GPT Image 2.5 Tested: Flare, Sunburst, and Astra as the Real Winner
OpenAI's GPT Image 2.5 fixes a couple of Image 2's nagging issues but plays it fairly safe on its own — the real upgrade is GPT-6 Astra riding along as the brain, driving conversational edits and research-backed prompts. I ran it through the standard tests, a four-angle grid matchup against Nano Banana, and a round of arguing with it over some Fury Road stills.
-
GPT-6 Astra Arrives With an AGI Claim, and World Models Have a Big Week
OpenAI says GPT-6 (codename Astra) opens the AGI era, backed by a 99.9% score on ARC-AGI-3 that comes with its own asterisk, plus new computer-use demos in Excel, Word, Blender, and Unreal Engine 5. Meanwhile, World Labs' Atlas, Runway's GWM Worlds 2, and a couple of MiniMax H3 hacks make it a loaded week for world models.
-
MiniMax H3 Max on fal: AI Video Faster Than You Can Watch It
fal's MiniMax H3 Max renders a 5-second clip with audio in under 3 seconds, and within days builders turned that into infinite AI TV channels, a 24/7 AI news stream, and prompt-driven games. I look at how it works, what people built with it over one weekend, what it costs, and whether it'll be open-sourced.
-
The Seedance 2.5 Prompt Guide: How Prompting Actually Changed
Seedance 2.5 is out globally, and its 80 pages of official prompting docs change the job: bracketed sections, numbered reference roles with declared exclusions, timeline scripts with end states, and four kinds of audio punctuation. Here's the full breakdown — plus the real production prompt from Death Walks Into A Bar and a free skill file for your LLM.
-
Flux 3 Is Here — and an AI Film Is Getting a Theatrical Release
Black Forest Labs finally shipped the video model teased since 2024: 20-second generations, up to 10 references, native audio, and open weights in the launch plan. Plus the landmark: Gossip Goblin's feature film is headed to actual theaters, and the workflow breakdown — Midjourney as the generator, Nano Banana as the Photoshop, first-frame image-to-video for camera control — is worth stealing.
-
Fine. Let's Talk About Higgsfield.
Higgsfield rewrote its Terms of Service twice in one week. Here's what the original July 23rd terms actually said — a perpetual output license, default-on training, a face license you could never take back — what the July 26th revision fixed, what it quietly walked back, and the one clause that survived word-for-word.
-
Real-Time AI Video Gets Cheap: Lucy 2.5 at Two Cents a Second
Decart's Lucy 2.5 restyles live video at 1080p/30 with near-zero latency, and API pricing around $0.02/second makes it cheap enough to actually build on — live previs might be the real use case. Plus: a new Seedance 2.5 sample, Blomkamp's Seedance short, SeedAudio timing control, Martini's camera motion feature, and depth maps as storyboards.
-
Seedream 5.0 Pro Tested: Layers, Blobby Faces, and Seedance 2.5 Leaks
ByteDance's Seedream 5.0 Pro generates in actual layers — up to 10 reference images, native 4K, and a layer editor on Dreamina that isn't fully baked yet. Plus: Seedance 2.5 slips to at least July 20 while the leaked clips keep getting better, LingBot-World 2.0 puts a real-time world model on Hugging Face, and Anam's CARA-4 gives AI voices a face.
-
Meta Finally Shipped an Image Model — and I Built You a Free Motion-Control Tool
Meta Muse Image is live and free at meta.ai — a thinking image model that argues with itself — with Muse Video on the way. The Banana survives, but it ranks higher than Meta's track record suggests. Plus: a free pose + depth tool I built for real motion control in your video-to-video generations.
-
Seedance 2.5's First Official Shots — and a Leaked 180-Second Beta
The first legit Seedance 2.5 samples are out on Volcano Engine Ark, and the story isn't prettier video — it's control: true video-to-video editing that changes one element while everything else stays locked, plus generations built from up to 50 references. And a prematurely-published Dreamina page teases a beta mode promising 180-second videos, with a 2.5 Fast and Mini reportedly following. Global release looks about two weeks out.
-
SeedAudio 1.0: Lock Your Dialogue Before You Pay for Seedance
ByteDance's new SeedAudio 1.0 isn't a music model — it's a cheap audio reference that locks your dialogue, performance, and ambiance before you spend on the expensive Seedance video pass. The audio-first workflow: running it on Fal, dodging context overload with Nano Banana Pro, timestamping your audio with an LLM, and what it costs. Plus camera control without Blender via Martini.
-
The Seedance News Everyone Missed: ByteDance's Volcano Ark Licenses Real Movies — Legally
Everyone broke down Seedance 2.5's specs. At the same event, ByteDance quietly launched Volcano Ark — a way to generate AI video from real films and a real movie star, with the rights-holder getting paid. The first partner is Stephen Chow's Bingo Group, it licenses real likeness and voice, and it's already live on Seedance 2.0.
-
Seedance 2.5 Is Official: 30-Second Shots, 50 References, Native 4K
ByteDance unveiled Seedance 2.5 at its Volcano Engine FORCE conference — native 4K, single-shot generation up to 30 seconds, and up to 50 multimodal references. Plus: Seedance 2.0's 4K is live right now (and it'll cost you), and a Kling research preview called OmniDirector hints at where camera control is headed.
-
Seedance and Kling Just Changed What "Fast" Can Do
The "fast, cheap, or good — pick two" rule cracked this week: Seedance 2.0 Mini, Kling 3 Turbo, two real-time models you can walk around inside, and a stealth Google image model the internet is calling Instant Ramen. What is actually usable now, and what is still a really good trailer.
-
They Killed Fable. Here's What You Need to Do Next
Anthropic's Claude Fable 5 was live for about three days before a government directive switched it off. What we actually know, what's still fog of war, and the lesson that outlasts the drama: build systems you own, not one-off projects.
-
Don't Buy an AI Film Studio. Build Yours.
The Paperclip Heart production breakdown — a Claude-powered production office that lives in one folder on your desktop, handles ref grids, scene boards, and asset tracking, and generates shots over MCP while you keep working.
-
The Cinematic Mega Prompt: From 2×2 Stills to a Moving Shot
The full prompt stack from the Cinematic Prompt Technique video — the Nano Banana Pro 2×2 template, the optional upscale pass, and the handheld video prompt. Copy, paste, direct.