Baidu Wizshot-ME: a 4000 character script in one pass
On November 12 at Baidu World 2026 the company introduced ERNIE Wizshot Micro-Expression, Wizshot-ME for short. The model makes video with talking characters, with text, image and voice generated together. It takes a script up to 4000 Chinese characters. Emotion gets described through a system of 6 main dimensions and 200 sub-dimensions. Voice runs in a single pass over the whole script; before, it was stitched from chunks of about 30 seconds. Businesses reach the model through the API on the Qianfan platform. Baidu claims a price 50% below the market, and the fast version turns out a minute of video in roughly a minute. No independent price or quality measurements existed on November 13, so the 50% figure stays a vendor promise.
Why this matters: a talking character on a long text stops being a pile of splices. For a series of explainer spots that means fewer cuts and fewer places where intonation breaks.
Baidu Yijing: from 0.2 yuan a second
On November 12 Baidu's digital human team, the Yijing project, showed its own video generation model at a price from 0.2 yuan a second. The fast version returns 10 seconds of video in 10 seconds, holds up to 8 to 9 facial micro-expressions a second, and answers in real time under 2 seconds. Alongside the model the team launched a studio for short animated dramas: script, storyboard, finished episodes and transfer to other markets.
Why this matters: at that price a draft episode stops being scarce. Scenes can run several times and be picked out of ten takes instead of one.

Google Flow: five tools built with artists
On November 12 the Google blog introduced five new Flow tools, each built together with artists. Character Poster handles a poster with a character. Character Arc shows one character at different ages. Pixel Machine goes into bitmap, ASCII, CRT, pixel art and glitch. Lucid Glass builds prisms, halos and refractions. Pocket Studio takes a single take shot on a phone and opens it into several angles, keeping the actor's performance.
Why this matters: Flow is growing the styles that used to be assembled in a separate editor. A poster, a character at three ages and a second angle from a phone are three routine jobs where handwork usually starts.
ElevenLabs Dub & Sync: dubbing and lips in one run
On November 12 Dub & Sync arrived in ElevenCreative. You upload a video, pick languages and get a dub with the lips fitted to the new voice track, all in one run. Lip sync used to be a separate step after the voice. The announcement carried no prices and no count of languages.
Why this matters: a foreign-language cut stops being a two-job budget line. For a client it is one item in the order, for a studio one handover.
Melius Ditto: up to 30 swaps per generation
On November 11 Melius CEO Joowon Kim wrote on X about Ditto, with documentation sitting next to it in the API. The tool swaps people, objects and backgrounds inside a finished clip, and motion and camera work stay as they were. One generation carries up to 30 swaps. It is in beta, entry through a waitlist, with an API and MCP.
Why this matters: a fix moves out of a reshoot and into post. The product in frame changes without a second shoot, and a version for another hero comes out of the same footage.
AI video news: DomoAI and Krea hand the whole project to an agent
On November 13 DomoAI announced InkoAI Agent through PR Newswire. From one sentence the agent builds a finished spot for TikTok, Shorts and Reels, shows a plan and waits for approval. Under the hood sit 15 image and video models, among them Seedance 2.5 and MiniMax H3. A day earlier Krea showed Krea Bots in a post from @krea_ai: the bots take on whole creative projects and tune themselves to a team's taste, still in beta.
Why this matters: an agent starts with a plan and works down to the shot. While it waits for approval a person stays the editor, and that is the last place in the chain where a decision gets made.
Diffusers 0.43, a GTA world model and Vidu Q4 Preview: short lines
Hugging Face shipped Diffusers 0.43.0 on November 12. The release added MiniMax-H3-Image, Pika H3 Flex with video and stereo sound in 6 steps, P-H3-Image, Kandinsky S2V 1.0 that builds a person from a photo and a voice recording in 5 second chunks, and PAN 2, a scene simulator at 29 billion parameters.
On November 11 and 12 lucidml published lucidml/gta on Hugging Face under Apache-2.0. It is an interactive world model at 630 million parameters, trained on 800 hours of GTA V. One starting image, WASD keys, frames generated one at a time and sent to a browser, output at 384×672. It is a research toy, and the rights to the training footage are left open in the model card.
Vidu opened Q4 Preview on November 11: faster generation, better prompt understanding, no numbers and no method in the announcement. On November 12 the company described an updated recaptioning. In the API, image to video runs 3 to 16 seconds, from 540p to 4K, with the video-with-sound option on by default. On November 13 Seedance 2.5 Flash turned up in Runway, and the details were in the November 12 roundup.
Why this matters: open releases and previews move at different speeds. Weights and a licence in Diffusers let you test a model on your own footage, while a preview without numbers gives you nothing to measure.
AI video news: how this lands in client work
For two days the market moved one way: a second gets cheaper and assembly collapses into a single run. Baidu already quotes a second in yuan at close to real time, ElevenLabs removed the separate lip sync step, Melius removed the reshoot for swapping an object. The November 12 roundup covered Seedance 2.5 Flash, Adobe and Luma, where the talk was clip length and image quality.
For a quote this changes the arithmetic. What counts is the number of iterations, since a cheap second over ten runs costs more than an average second over two. AI video production at volume runs on that logic, and so do vertical video ads, where a set holds more versions than ideas.
A channel works the other way. Running an AI YouTube channel means shipping episodes for months, and there the price of a second matters less than the same face and the same voice from episode to episode. Today's releases promise exactly that consistency, and it still has to be checked on your own scenes.




