AI video news, November 14: Duo v1, Pine, YouArt, Bloomway

A Saturday, so the news is thin and almost all of it is about assembly. inclusionAI released Duo v1, which answers with video and voice in real time and lets itself be interrupted, Pine opened a desktop studio for micro-dramas, YouArt put an agent inside Premiere and After Effects, Bloomway turns a sentence into a rough 3D scene. Then a rumor about LoRA on Gemini Omni, a Higgsfield contest with a $1,000,000 fund and cuts at Grok Imagine.

Blank storyboard cards ride a conveyor while one stamp marks a single approved card
01

Duo v1 from inclusionAI: a model you can interrupt

On November 13 inclusionAI published Duo v1: code on GitHub, weights on Hugging Face, a paper on arXiv. It is a model for live video conversation, full duplex: it listens and watches the person it talks to while answering with video and voice at the same time, and it can be interrupted mid-sentence. Two parts sit inside. Intent decides when to speak and what to say, Realization draws the video with sound. Duo 1.0 is the base version, distilled to 4 steps in FP8, and Duo 1.2 sits next to it. The code is Apache-2.0, the weights were initialised from MiniMax H3. This is a preview for researchers: running it takes two CUDA GPUs. According to AI Weekly, replies arrive in chunks of 0.71 s, 93.3% of criteria passed across 30 test conversations, and the first reaction lands in 330 to 650 ms on 8 GPUs. The numbers come from the authors' paper, with no independent measurements yet.

Why this matters: a conversation with a digital person stops being a scripted recording. Support and consultation in frame get a format where the viewer cuts in and gets an answer in voice.

02

Pine Drama Studio: agents build the series, a person approves

On November 13 Pine AI opened Pine Drama Studio through PR Newswire and the PineComputer/pine-drama-studio repository, a desktop studio under Apache 2.0 for AI micro-dramas. Input is a script and one instruction. Agents then build the story bible, the design of characters, locations and props for the whole series, the storyboard, generate shots, check takes, redo the ones that drifted and cut the episodes. A person approves the design, the storyboard, the takes and the final cut. Generation itself runs on Pine Computer's cloud, and that is the paid part.

Why this matters: a series stops stalling on a character checked by hand on every shot. The bottleneck moves to approvals.

Four pipeline stations: live video dialogue, series frames, agent inside an editing frame, 3D camera
03

YouArt Desktop: an agent inside Premiere and After Effects

On November 13 YouArt, a Y Combinator startup, shipped a desktop app for Mac and PC through GlobeNewswire. With the user's permission the agent works inside After Effects, Premiere Pro, DaVinci Resolve, CapCut and Blender: it opens menus, clicks, sends commands. A client note that arrives in Slack or Gmail gets pinned as a comment to the right shot. Your own API keys for Higgsfield, fal and BytePlus connect directly, and the routing looks for the cheaper provider. Seedance 2.5 at 720p costs $0.208 a second, and $0.138 a second on the annual Max plan.

Why this matters: a client note reaches the timeline without a retelling. A Slack comment turns into a task on one shot, and carrying notes over usually eats a lot of time.

04

Bloomway 3D Director: a rough 3D scene from a sentence

On November 13 Bloomway AI showed 3D Director Studio through GlobeNewswire. You describe a world in a sentence and get a simplified 3D scene where action and camera movement are set in words, the announcement uses "She steps outside. The camera slowly pulls back." Then you see an animated previs, a rough 3D blockout with stand-ins for people, and that same previs becomes the reference for the final generation. Camera position and path, focal length, shot size and angle can be checked and fixed before rendering.

Why this matters: camera movement stops being a lottery inside a prompt. A director places it in the blockout and the model gets a ready blueprint for the frame.

A level balance scale: ten small cubes in one pan, two large cubes in the other
05

AI video news: HyperFrames built an agent on DeepSeek Harness

On November 13 the HyperFrames blog ran a post titled "Our harness beats Claude Code: here's how". HyperFrames Studio, the free app where video gets assembled in code, now carries its own agent on DeepSeek Harness. On their test of 17 video tasks with the same model Claude Code uses, the run came out about 30% cheaper, the model took about 35% fewer turns, 78% of tasks needed less effort, and remixing a reference cost 40% less. Quality in a blind pairwise comparison was 41% against 41%, and a model did the judging, with no human panel.

Why this matters: the harness around a model moves the price as much as the choice of model. Where assembly runs in code, a 30% price gap shows up at series scale.

06

Rumor: a LoRA spec for Gemini Omni turned up

On November 13 US time the tracker bot @aitrackerbot wrote on X that the Vertex AI API, branch v1beta1, carries a spec called "Omni LoRA tuning". According to the post it is video model fine-tuning on your own JSONL set for text to video and image to video, with adapter sizes of 8, 16 or 32. Google announced nothing of the kind, and as of November 14 the documentation holds nothing. For now the claim rests on a single post.

Why this matters: if the spec is real, a series style can live in your own adapter instead of a long prompt repeated in every shot. The test has to run on your own footage once access arrives.

07

Higgsfield, K2V and Grok Imagine: short lines

On November 13 the Higgsfield changelog opened Talent Contest, an AI film competition with a $1,000,000 prize fund and 20 places. Films run 3 to 7 minutes, any theme, built in Cinema Studio, English subtitles. Entry is free, 18+, solo or a team of up to 4.

Cambridge and Toyota Motor Europe published K2V on arXiv on November 12: Generative Keyframing For Physically Plausible Video Generation. A language model plans the physics of a scene and produces keyframes, and the video model fills the gaps. The authors report the best result on the Physics-IQ Verified and PhyGenBench tests, and their table puts a clip at $0.47 to $0.59.

On November 13 The Information reported that the AI unit of SpaceX, the former xAI, cut about 60 people from the Grok Imagine team, which makes video and images, over a week. A handful stayed to keep the product running, and code tools now hold priority. No official statement exists, and this is a press report.

Why this matters: three lines show where the money sits. Competitions pay for a finished film, researchers price a clip, and a large player shrinks video down to maintenance. For a studio the takeaway is simple: do not tie a pipeline to a single model.

08

AI video news: what this changes in an order

Duo v1 opens a conversation in real time, Pine builds a series, YouArt sits inside someone else's editing app, Bloomway places the camera in rough 3D. They share one direction: work moves out of prompt typing and into approvals and notes. The November 13 roundup covered Wizshot-ME, the Flow tools and Dub & Sync, where the talk was the price of a second and the length of a shot.

For a quote this changes one number. A cheap second over ten runs costs more than an average second over two, so AI video production bills iterations and renders as separate lines. Vertical video ads ship in versions, and there camera notes matter more than resolution.

An AI YouTube channel works the other way: episodes run for months, and the same face and voice matter more than the price of a second. Today's releases promise that consistency through a story bible and a previs, and it still has to be checked on your own scenes.

frequently asked

01What does full duplex mean here
Duo v1 listens and watches while answering with video and voice at the same time, and it can be interrupted mid-sentence. One part decides when to speak, the other draws the video with sound.
02What does Pine Drama Studio do
It builds an AI micro-drama from a script and one instruction: story bible, character and location design, storyboard, shots, take checks and the edit. A person approves the design, storyboard, takes and final cut.
03Does YouArt work inside Premiere and After Effects
Yes, with the user's permission the agent works inside After Effects, Premiere Pro, DaVinci Resolve, CapCut and Blender through menus and clicks. Your own Higgsfield, fal and BytePlus keys connect directly.
04How much is a second of Seedance 2.5 in YouArt
$0.208 a second at 720p, and $0.138 on the annual Max plan. The numbers come from the company, with no outside check.
05Is the Gemini Omni LoRA available now
No. As of November 14 it is a rumor: a tracker bot spotted an "Omni LoRA tuning" spec in the Vertex AI API, Google has announced nothing, and the documentation says nothing.
06What is the prize fund at the Higgsfield contest
$1,000,000 across 20 places. Films run 3 to 7 minutes, built in Cinema Studio, English subtitles, free entry from 18.