标签：#Video

Several Hollywood celebs are ditching the massive eight-figure checks and exotic movie sets for a rising format: microdramas.

Grand Theft Auto VI is nigh. Here’s what the developer revealed about its highly anticipated game.

TL;DR We're building a caption evaluation harness that scores a WebVTT file on four axes instead of one: word error rate under a fixed normalizer, missed entity rate on domain terms, median cue timing offset, and reading rate in characters per second. Python 3.12, jiwer , whisper_normalizer , webvtt-py . Run it on every model or vendor change. A caption file can score 96% accurate and still be unusable. WER counts substitutions, insertions and deletions and weighs each one the same, so "fifteen milligrams" becoming "fifty milligrams" costs exactly as much as "the" becoming "a". It also throws away every timestamp before it starts, which means synchronization and readability are invisible to it. Let's measure the other three things. 0. Setup 🛠️ python3 -m venv .venv && source .venv/bin/activate pip install jiwer whisper_normalizer webvtt-py $ pip list | grep -Ei 'jiwer|whisper|webvtt' jiwer <your version> webvtt-py <your version> whisper-normalizer <your version> Pin whatever you install, and pin it in CI. The APIs below move between majors, which is exactly why the next tip exists. 💡 Tip: jiwer.compute_measures() is gone in recent versions. It is jiwer.process_words() now, and it returns a WordOutput dataclass. Most blog posts you will find still use the old name. 1. Parse the VTT into text plus timings # captions.py from dataclasses import dataclass import webvtt @dataclass class Cue : start : float end : float text : str @property def duration ( self ) -> float : return self . end - self . start @property def lines ( self ) -> list [ str ]: return self . text . split ( " \n " ) @property def flat ( self ) -> str : return " " . join ( l . strip () for l in self . lines ) @property def chars_per_second ( self ) -> float : return len ( self . flat ) / self . duration if self . duration > 0 else float ( " inf " ) def _to_seconds ( ts : str ) -> float : h , m , s = ts . split ( " : " ) return int ( h ) * 3600 + int ( m ) * 60 + float ( s ) def load_vtt ( path : str ) -

TL;DR -c copy can only cut on keyframes, so your 12.4s trim starts wherever the last keyframe was. We'll build a smart-trim script that probes keyframe positions with ffprobe , re-encodes only the head and tail fragments, stream copies everything between them, and concatenates the three. Frame accurate output, encoding cost proportional to two GOPs instead of the whole file. Tested with FFmpeg 9.0 "Lei" (released 2026-08-04) and Node 22.x. The JS is ESM, so put "type": "module" in your package.json before running any of it. Everything here also works on FFmpeg 7.x and 8.x; nothing we use is new. The problem, in two commands 🎬 # fast, and wrong ffmpeg -ss 12.4 -i input.mp4 -t 20 -c copy fast.mp4 ffprobe -v error -show_entries format = start_time,duration -of default = nw = 1 fast.mp4 # start_time=0.000000 # duration=20.388000 <- we asked for 20, starting at 12.4 The clip is long by the distance from our requested start back to the previous keyframe, and every frame in it is shifted earlier than the user asked for. Stream copy moves compressed packets without decoding them. Most frames in a compressed stream only describe the difference from their neighbors, so the only place you can start is a keyframe. FFmpeg snaps back to the nearest preceding one, and your clip starts early. # accurate, and slow on a long source ffmpeg -ss 12.4 -i input.mp4 -t 20 -c :v libx264 -crf 20 -c :a aac slow.mp4 We want the accuracy of the second and roughly the cost of the first. 1. Look at your keyframes first Before writing any code, find out how bad the problem is for your content: ffprobe -v error -select_streams v:0 \ -show_entries packet = pts_time,flags \ -of csv = print_section = 0 input.mp4 | grep 'K' | head -20 0.000000,K__ 2.002000,K__ 4.004000,K__ 6.006000,K__ Two second GOPs here, so worst-case error is about two seconds. Screen recorders and some camera output emit keyframes on scene change only, and there the gaps can be 30 seconds or more. That distribution is the real spe

Variant Multiplier already let an editor swap one section of a winning ad and keep the rest. The next request from a real production job — replacing product SL-603 with SL-808, a different hearing-aid SKU, across an entire finished ad — was a different shape of problem. It's not "change one section," it's "change every mention of the product, everywhere it appears, while keeping literally everything else the same." Two direct quotes from the editor drove the whole five-PR arc: the transcript editing was too rigid for word-by-word changes, and separately, "the music, voice, etc. should retain the same, we should keep the quality the same, and not make it do a lot of changes." If a re-render can degrade something the editor explicitly asked to keep untouched, the render path is wrong for the job — no matter how good the model is. The cheap fix first: let editors actually edit PR #67 shipped before any product-swap work started, because it was the cheap, high-value half of the same feedback: "I am just able to select word by word here but I am not really able to change the whole sentence a lot easier," and separately, "I'm able to double click on these words and then just type it in." Both were UI gaps in the transcript editor, not pipeline gaps — selecting by sentence or scene instead of only by word, and retyping a line verbatim instead of only substituting individual words. Shipping this first, standalone, meant the harder product-swap work that followed didn't also have to carry an unrelated UX fix in its diff. A product catalog the tool never had PR #69, stacked directly on top of the transcript work, is pure groundwork with no user-visible feature of its own: a product catalog, because Variant Multiplier had no concept of "a product" at all before this. The editor's own framing made the requirement explicit: "have a product selection right here, for Pro Bluetooth, for [the other SKU], and maybe other tons of products" going forward. The catalog data itself is mai

On the AI video ad platform I work on, every scene goes through the same painful loop: write a prompt, send it to an AI video model provider, wait two minutes, open the result, squint at the frame, and decide what went wrong. Camera too wide. Product missing from the hero shot. Color palette drifted warm when the brand brief says cool neutrals. Avatar looks like a different person than scene three. That loop was manual, slow, and expensive. Each regeneration burns GPU credits. Operators were becoming prompt engineers by accident — and still missing subtle failures until stitch time, when fixing scene four means re-rendering everything downstream. The insight behind vision-in-the-loop prompt authoring is simple: the model that wrote the prompt can also look at its own output and rewrite the prompt with surgical fixes. Not a full replan — a per-scene correction grounded in the actual generated frame, not the operator's memory of what they hoped would appear. The manual loop we were trying to kill Before this work shipped, the swipe iteration flow looked like this: Plan — Claude generates a scene-by-scene script with visual prompts Generate — each scene renders independently through an AI video model provider Review — operator opens the portal, compares frames to the reference ad Rewrite — operator edits prompts in a text field, often guessing at what the model misread Regenerate — repeat until acceptable or budget exhausted Steps three and four are where throughput dies. An experienced operator can spot "product not visible" in three seconds, but translating that into prompt language — "medium close-up, product centered in lower third, shallow depth of field" — takes another minute per scene. Multiply by twelve scenes and three swipe iterations, and a single ad creative consumes an hour of human attention that should be spent on brand strategy, not frame inspection. The generated frame is ground truth. The original prompt is a hypothesis. Vision-in-the-loop closes the

TL;DR In generative video pipelines, running cheap low-step sketches to pick parameters sounds like free optimization. But when prompts go out-of-distribution, surrogate scorers return noise, turning a \$0.002 check into a bad decision that triggers a \$15 compounding failure. Here's why skipping the cheap step is sometimes the cheapest option. Three numbers run Scenematic's generation loop. A think-frame costs \$0.002. A full render costs \$0.50. A bad scene that slips through and gets built on costs about \$15.50, because the scene chain compounds it before anyone looks. The constant in lib/generation-loop.ts carries the arithmetic in a comment: 15.502, // CALIBRATION_TARGET: 0.002 + 0.50 + 15.00 . Most of the pipeline exists to keep spend at the cheap end of that ladder. One module decides when the cheap step should be skipped entirely. A hundred-contract baseline then put numbers on how often that decision was wrong. 1. The rehearsal lib/think-frames.ts generates quick, low-inference-step sketches before committing to a full-quality keyframe. The file header credits DeepGen's think tokens as the inspiration. Each sketch tries a different preservation focus, character, environment, mood, composition, or atmosphere, with its own image-to-image strength and seed. The reward mixer scores the batch and the winner's parameters go to the full render. The economics only work if those scores mean something. That assumption fails quietly, and it fails hardest on the prompts where a rehearsal looks most useful. 2. Where the scores stop meaning anything Scoring a sketch of A detective leans forward across a metal table, interrogating a nervous suspect under fluorescent lights works fine. The scorer has seen a thousand shots like it. Scoring A sentient equation writes itself across a blackboard that extends infinitely in all dimensions does not fail loudly. It returns a number, and the number is noise. Both prompts are verbatim from the baseline harness. lib/ood-detector.ts

Investors want founders who understand the financial reality of their business. Messy data, misunderstood metrics, or waiting until you’re nearly out of cash to start fundraising can cost founders leverage, valuation, and even a term sheet. In this episode of Build Mode, host Isabelle Johannessen sits down with Sasha Orloff, founder and CEO of Puzzle […]

OpenAI presented details of its AI’s model’s cyberattack on Hugging Face at Black Hat last week. Simon Willison details the timeline. It’s really interesting to read through—and really impressive cyberoffense work.

Live video looks simple until you build it. Then you discover that "low latency" means five different things, that your CDN and your latency target are fighting each other, and that the box which handled ten viewers falls over at ten thousand for reasons nobody warned you about. This is the guide I wish existed when I started. No vendor talk, just how the pieces fit. 1. Ingest and delivery are separate decisions The single most common mistake is treating "streaming protocol" as one choice. It is two. Ingest is getting video from a camera, encoder or browser into your server. Delivery is getting it from your server to viewers. They have different constraints and you almost never use the same protocol for both. A typical stack ingests over RTMP or SRT and delivers over HLS. Another ingests WebRTC and delivers WebRTC. Mixing is normal and expected. Once you separate them, most of the confusion disappears. 2. The ingest protocols RTMP is old, TCP-based, and still everywhere. Every encoder speaks it, OBS defaults to it, and it just works. Latency is typically 2 to 5 seconds. Classic RTMP is limited to H.264 and AAC, though the Enhanced RTMP spec has added HEVC and AV1. Being TCP, it degrades badly on lossy networks: packet loss becomes head-of-line blocking, and your stream stalls instead of gracefully dropping quality. SRT is the answer to that. UDP-based with its own retransmission layer (ARQ), a configurable latency buffer, and built-in AES encryption. It is designed for pushing broadcast-quality video across the public internet, which is exactly where RTMP struggles. If your source is on a flaky connection, a 4G link, or a different continent, SRT is usually the right call. # Publishing over SRT with ffmpeg ffmpeg -re -i input.mp4 -c copy -f mpegts \ "srt://your-server:4200?streamid=live/stream1" RTSP is what IP cameras speak. If you are pulling from surveillance hardware, you are pulling RTSP whether you like it or not. WHIP (WebRTC-HTTP Ingestion Protocol) is the n

Every video card on our category grids was hotlinking a 1280x720 JPEG from a third-party CDN and then letting CSS scale it down to about 320 device-independent pixels. That is roughly 90 KB of wasted transfer per card, 24 cards per page, across eight regional page variants that each carry their own cache key. Mobile LCP on the busiest category pages sat at 4.1s, and the largest single contributor was an image we did not host, could not resize, and could not re-encode to WebP. The fix was not clever CSS. It was owning the frame. We built a small Go service that takes a source video (a partner preview MP4, or a poster frame that arrives at the wrong dimensions), pulls a representative frame with FFmpeg, encodes it at three widths in WebP, and writes the result to a content-addressed path the front end links directly. That service now feeds the same multi-region cron that runs TrendVidStream , and the generated files ride the same FTP mirror as the rest of the deploy. What follows is the part that actually mattered: the FFmpeg invocations, the Go concurrency model that keeps a 2-core build box from melting, and how a stateless Go daemon hands work to a PHP 8.4 + SQLite front end that cannot run a daemon at all. Why this is not a PHP job Our front end is PHP 8.4 on LiteSpeed shared hosting with SQLite (FTS5 for search) as the only datastore. It is a genuinely good fit for a read-heavy discovery site: no database server to babysit, page cache on disk, cron jobs pulling regional feeds every 2-7 hours depending on the site. It is a terrible fit for thumbnail extraction: Shared hosting caps max_execution_time at 180s. A cold FFmpeg decode of a 4-minute 1080p preview can burn 20-40s. Do 200 of them in one cron tick and you are wearing a hard timeout. shell_exec is frequently disabled, and when it is not, you get one process per request with no way to bound total concurrency. There is no shared memory between PHP requests, so two cron ticks racing on the same video ID will ha

The universe may never tell you if your choices mattered. Owlcat’s Osiris Reborn might not either.

Higgsfield, founded by former Snap exec Alex Mashrabov, lets users create AI images and videos.

The change comes a year after YouTube applied the same approach to counting views on Shorts videos.

Fascinating video about searching for life undersea. The video basically makes the point that our bright white searchlights are scaring everything away, and that red light is more neutral. That, plus bait to attract sea creatures, is teaching us a lot about what’s going on down there. Lots of footage of giant squid, and speculation about the colossal squid. Worth watching. As usual, you can also use this squid post to talk about the security stories in the news that I haven’t covered. Blog moderation policy.

Investors are still waiting for their share of the $250 million windfall, and VideoVerse co-founder Vinayak Shrivastav is now at the center of multiple legal cases.

Researchers say it took fewer than 20 prompts for a public AI tool to find a flaw (now fixed) allowing anyone on a Zoom call to hijack another participants’ device.

Nice video of the Arctic bobtail squid. As usual, you can also use this squid post to talk about the security stories in the news that I haven’t covered. Blog moderation policy.

Intro Day 20! I lined up 10 AIs that turn a single photo into a few seconds of video. Half ran locally on my DGX Spark, half in the cloud 🐱 What I used: DGX Spark (LTX-2.3 / Wan 2.2) / 8 cloud models via fal.ai / ComfyUI / ffmpeg The setup Item Value Input One identical photo (my cat on a desk) Length 6 seconds Settings Identical The only variable The prompt Easy prompt The cat looks at the camera and meows once. It opens its mouth, meows, then closes it. Its tail flicks and its ears twitch. Hard prompt The cat stands upright on its hind legs in a kitchen, wearing a small apron, holding a knife in its front paws and chopping vegetables on a cutting board. Steam rises from a pot behind it. Please, just watch it Some of the cats came out with very long legs. Anyway. First half is the easy prompt, second half the hard one. On the easy prompt, local and cloud were a fair match . On the hard one... cloud, I think...! Three rankings below. Ranking 1: Time Time per 6-second clip on the hard prompt. Rank Model Where Time 🥇 LTX-2.3 Cloud 41s 🥈 Wan 2.7 Cloud 92s 🥉 Happy Horse 1.1 Cloud 97s 4 Veo 3.1 Cloud 128s 5 Kling 3 Pro Cloud 205s 6 Seedance 2.0 Cloud 210s 7 LTX-2.3 Local 315s 8 Wan 2.2 Local 651s 9 daVinci-MagiHuman Cloud 710s 10 HunyuanVideo 1.5 Cloud 796s A 19x spread. Look at 1st and 7th. Same model, LTX-2.3 , nearly the same resolution. The only difference is where it ran — 7.6x . Local setup DGX Spark (GB10, 128GB unified memory, ~273GB/s). ComfyUI headless, workflows over its API. LTX-2.3 is distilled fp8 at 8 steps. At 1088×1920 peak memory hit 77.8GB, about 60% of 128GB. That was the ceiling. Dropping to 512×768 finishes in 70s, but with one-fifth the pixels. Wan 2.2 is I2V-A14B fp8, 20 steps, 480×640. Higher resolution does not finish in reasonable time. Ranking 2: Cost Rank Model Per 6 seconds 🥇 Local Electricity only 🥈 LTX-2.3 (cloud) $0.36 🥉 Wan 2.7 $0.90 4 Kling 3 Pro $1.01 5 Happy Horse 1.1 $1.08 6 Veo 3.1 $2.40 7 Seedance 2.0 $4.09 — HunyuanVideo / MagiHum

JioHotstar explains the distributed architecture behind its real-time ad request workflow, covering ad decisioning, waterfall tiering, pacing algorithms, latency optimization, and service coordination required to select and deliver personalized advertisements during streaming playback at scale. By Leela Kumili