AI Image, Video and Voice

Oct 28

AssemblyAI, known for speech-to-text, has launched AssemblyAI-TTS-1 for voice agents. It prices at $15 per million characters with no concurrency tiers. Most numbers come from AssemblyAI itself; here is what is independently ranked and what you should test.

Oct 28

Lightricks has published LTX-2-Marking on Hugging Face, a weights pack that lets the open LTX-2 pipelines embed an invisible PixelSeal mark in video and an AudioSeal mark in sound. It is opt-in, not cryptographic, and removable by a determined attacker. Here is how it works and what it means.

Oct 28

Rime launched Prism on October 28, 2026: a speech-to-speech model that writes what it will say before it speaks, so guardrails can read and rewrite it. Secondary coverage quotes a 300ms response; Rime's own numbers are 0.57 seconds and about 500ms.

Oct 27

Microsoft AI previewed MAI-Image-2.7 on October 27, 2026, edit-only for now in the MAI Playground, with a reported No. 2 on Artificial Analysis image editing behind OpenAI. Here is the claim, the context from earlier MAI-Image releases, and the gaps.

Oct 26

Oct 24

Kromoss is a community full fine-tune of OpenMOSS's MOSS-Image-Base from lodestones, trained on 704k images and released under Apache 2.0 with two 8-step turbo LoRAs. It has been trending on Hugging Face. Here is how to run it and how far its one benchmark goes.

Oct 23

On October 23, 2026 Block announced Dessa, a research-preview voice model it says was trained from scratch on nothing but human conversation audio, listens and talks at the same time, and runs locally on a MacBook Pro M4 Max. There are demos and a waitlist, but no paper, weights or benchmarks yet.

Oct 22

On October 22, 2026 Canva released Craft, its first in-house image generation model. It plans the layout before it draws, is trained on human-designed templates rather than photos, and ships to every Canva user through Dream Lab and the Canva API. Here is what is confirmed, what is not, and how it fits Canva's wider AI push.

Oct 22

The FastVideo project, with Nuva Lab and NVIDIA, released FastH3 Omni: an eight-step distilled version of MiniMax-H3 that takes a prompt plus up to twelve image, video and audio references and returns video with synchronized stereo audio. Weights and training code are open, under the MiniMax-H3 community license.

Oct 22

Gradium released the Phonon family of tiny text-to-speech models. Phonon-31m is Apache 2.0, about 65 MB quantized, and streams speech from a browser tab or a laptop CPU. The larger 40m models are noncommercial.

Oct 20

Bria released Fibo 2.0, an 8B open-weights model that generates and edits images from structured JSON captions. Artificial Analysis lists it a hair behind Qwen-Image-2.1 for open text-to-image and first for open editing, but the weights are non-commercial and the $32 per 1,000 images figure is a hosted-API claim.

Oct 20

ElevenLabs released Eleven Realtime on October 20, 2026: two end-to-end speech-to-speech models, v1 Mini for natural conversation and v1 Max for tool-calling tasks. Here is the benchmark table, how the OpenAI-compatible API works, which numbers are the vendor's own, and what to test before you switch.

Oct 20

Lightricks has put LTX-2.5 on Hugging Face: a 22B-parameter open-weight model that generates synchronized video and audio, now with native multishot generation. Here is what the model card says, what the license allows, and how to run it locally.

Oct 20

Ask an AI image editor for five edits in a row and the picture usually decays: textures smear, faces change, earlier edits vanish. MT Forcing, a new paper with released code and checkpoints, trains the model on its own multi-turn rollouts to stop that drift. Here is what it does and what the numbers do and do not show.

Oct 20

OpenFilm, an MIT-licensed repo that gained 1,338 stars in about two weeks, turns any coding agent into a video agent: a film is a folder of web pages with a frame(t) function, and a local Studio lets you change anything the agent made. Here is how it works, what it costs, and where it breaks.

Oct 20

Every latent diffusion image model sits on top of a VAE tokenizer, and a good one has to be sharp and easy for a diffusion model to learn. RoAR-VAE, from Tsinghua University and Microsoft Research, argues that a simple principle called robust prediction gets both. The code is not out yet, so here is what is claimed and what to check.

Oct 20

Mistral AI released Voxtral 3 on October 20, 2026: speech-to-text models, a new expressive text-to-speech model, a Pro Voice adaptation service and a voice agent SDK built on LiveKit. Here is what shipped, the prices, and which benchmark numbers are Mistral's own.

Oct 19

fal Research, working with Black Forest Labs, released FLUX 3 Max, a post-trained FLUX 3 Video variant. fal says it ties for first in a preliminary Artificial Analysis image-to-video evaluation and renders 10 seconds in about 11 seconds. Here is what the model page confirms and what is still the vendor talking.

Oct 19

Kling AI has released version 4.0 of its video generator. We list what the company says it ships today, what is still marked coming soon, and how it stacks up in early side-by-sides with Wan 3.0 and Seedance 2.5.

Oct 19

A paper trending on Hugging Face proposes Speculative Sampling, a training-free serving trick that keeps big bidirectional attention windows but delivers video incrementally. The authors report about 1.4 seconds from a new input to the first responding frame on MiniMax H3, and the code is open.

Oct 17

OpenMOSS published MOSS-Image-Base and MOSS-Image-SFT on October 16, 2026: a 5B text-to-image model under Apache 2.0 that renders text, accepts JSON layouts with bounding boxes, and takes prompts up to 4K tokens. Here is how to run it and what is not yet proven.

Oct 16

Elorian, the Palo Alto startup founded by ex-DeepMind researcher Andrew Dai, shipped its first model, Glimpse, as a research preview. It says the model beats GPT-6 Astra on chart reasoning and short video captioning. Here is how it works, what it costs, and which claims are still unchecked.

Oct 16

Posts from OpenRouter and Higgsfield say Black Forest Labs' video upscaler now costs about half as much for 4K output, and that FLUX 3 Video reaches 2K and 4K. We checked the numbers against OpenRouter's listing, the August launch prices and BFL's pricing page.

Oct 16

At a Fortune summit, a Yum Brands executive said the Taco Bell drive-thru bot is increasing retention because it makes jobs easier. The claim traces back to a vendor press release with no published figures. Here is what is confirmed and what is not.

Oct 15

Black Forest Labs has released FLUX 3 Image Klein: 5.6B parameters, open weights, native 4K, multi-reference editing and bounding-box control, with an INT8 build that fits an 8GB RTX card. Here is what is claimed, what the license allows and how to run it.

Oct 15

Krea launched Krea 3, an image editing model it says finishes most 2K edits in under 8 seconds at $0.06 per image, with up to 6 reference images. The speed and price comparisons are Krea's own, so here is what is claimed and how to check it.

Oct 14

Deezer says a new toggle will let listeners remove AI-generated tracks, albums and related artists from their experience, due before the end of 2026. We cover what is confirmed, the numbers behind it (over half of daily uploads are AI), how Spotify, Bandcamp, Tidal and Apple Music differ, and what the filter cannot do.

Oct 14

Mirelo launched SFX 2.0 on October 14, 2026, a video-to-sound model that keeps the dialogue, removes unwanted music and generates synced effects, foley and ambience for videos up to 10 minutes. We go through the vendor's blind-test table, the three model variants, where it runs and what to check before using it in a pipeline.

Oct 14

On October 14, 2026 YouTube began tagging AI-made songs in the YouTube Music app, with the main YouTube app to follow. We explain the labels, the partner disclosure scheme, what is not penalized, and how Spotify, Deezer, Apple Music and TIDAL differ.

Oct 13

On October 13, 2026, Baroness Campbell of Surbiton became the first person to use an AI-generated voice to address the UK Parliament, using an ElevenLabs clone of her own voice. explainx.ai covers how it was made, the approval, the quotes and what it signals for assistive AI.

Oct 11

A BBC analysis by Ros Atkins argues that a flood of AI content is making obviously human-made art more valuable, but that the premium may not last. We separate the Deezer numbers, the expert claims and the Hacker News reaction.

Oct 9

Zack London, known on YouTube as Gossip Goblin, made an AI feature that the filmmakers say is the first AI film to get an MPA rating. It is rated R, moves to December 4, and will go for Best Animated Feature. Here is what trade press confirms and what rests on the producers’ word.

Oct 9

Sperid Labs released Iris-3B on October 8, 2026, a 3B text-to-image diffusion transformer that outputs every pixel directly, with no VAE and no latent space. The paper reports the model matches Qwen-Image on OneIG. It also reports that pixel space did not beat latent models on depth or restoration.

Oct 9

Voyager launched on October 8, 2026 as an "open harness" for creative work. The founder calls it the Codex for creative work. We read the launch thread, the FAQ and the pricing page. One surprise: the app is not open source.

Oct 6

Kandinsky Lab open-sourced Kandinsky 6.0 Video on October 6, 2026: two diffusion models that generate five-second clips with synchronized 44 kHz audio, released with code and weights under the MIT license. This post covers the architecture, the claimed results, the GPU requirements, and where it fits next to other open video models.

Sep 29

On September 24, 2026, Dev Mode Discord screenshots showed VIDEO 4.0 Preview and VIDEO 4.0 Flash Preview inside Kling's creator app, with a duration slider from 3 to 30 seconds. Kuaishou has not announced the model, published a model card, or listed 4.0 in the API changelog. This news post separates leaked UI from what you can ship on today.

Sep 26

Sep 26

Sep 24

Fish Audio introduced Drama 3, a preview text-to-speech model it calls the most controllable ever: describe tone and character in simple language, change voice mid-sentence, render multi-character scenes and regenerate a single word. Access is gated, pricing is unpublished. Here is what is confirmed, what to test, and how it compares.

Sep 20

A viral blog post and its 750-comment Hacker News argument prove that "AI slop" posters aren't a model limitation — they're what happens when nobody tells the model which design language to use. The same fix applies to AI-generated code and prose.

Sep 17

Sep 15

Sep 10

Sep 10

Sep 9

Sep 7

Sep 3

Sep 2

Sep 2

Sep 2

Sep 1

Aug 31

Aug 31

Aug 31

Aug 29

Aug 29

Aug 29

Aug 28

Aug 28

Aug 28

Aug 27

Aug 26

Aug 25

Aug 22

Aug 19

Aug 18

Aug 18

Aug 18

Aug 9

Aug 5

Aug 1

Jul 31

Jul 31

Jul 30

Jul 29

Jul 26

Jul 23

Jul 19

Jul 19

Jul 15

Jul 10

Jul 9

Jul 8

Jul 8

Jul 7

Jul 4

Jul 3

Jun 27

Jun 27

Jun 25

Jun 23

Jun 23

Jun 22

Jun 21

Jun 21

Jun 20

Jun 19

Jun 18

Jun 17

Jun 16

Jun 13

Jun 4

May 31

May 27

May 26

May 24

May 22

May 21

May 20

May 4

Apr 25

Apr 13