Skip to content
Date and source

Dates follow the news timeline in Beijing time and include both endpoints.

Oct 1Thu
  1. Theoretically Media blog62

    Early Hands-On with Kling 4, Plus a Look at Gemini 4 Argon, Ideogram 4.5, and Eleven v4

    The author conducted hands-on testing of the Kling 4 early access version, which currently only offers the Flash model with 720p output, available to Ultra subscribers on annual plans; larger models are planned for October. Testing shows that Kling 4 appears to have resolved the lip-sync drift issue from 3.0, supports up to ten keyframe references, but performances are somewhat flat.

  2. Suno 官方博客62

    Suno releases the Speech audio model beta, capable of generating vocals and music simultaneously

    Suno released Speech beta, its first audio model that generates human voice and music together as a single complete track. Users input an idea, poem, or existing text, then describe the desired vocal tone and music style, and the system generates voice content with original background music. The model is now open to all users in beta; it had previously been tested with a small group of users for one month. The official statement notes that British accents occasionally drift toward Australian, and other effects still need improvement.

    Why it matters: Suno synthesizes voice and background music into a single track, allowing creators who do short-drama narration or dubbing to evaluate a new production approach.

  3. Runway48

    Runway releases world interface model Solaris

    At today's AI Summit, Runway CTO Kamil Sindi further introduced our first world interface model, Solaris. A new type of operating system that builds itself as you interact. Every frame is generated in real time based on your intent. In the keynote, Kamil shared a clear vision: in the future, all software will be generated. Without a single line of code. Learn more about Solaris: https://runway.com/news/research/introducing-solaris

Sep 30Wed
  1. LTX-2 官方发布17

    LTX-2 v1.4.1 released

    LTX-2 released v1.4.1, modifying the HDR video encoding process to output decoded data blocks directly as EXR and HLG, no longer retaining the full clip in memory. This version was released by lt-research-bot on September 30 and includes 1 commit to the main branch.

  2. 可灵 Kling60

    Kling showcases a short film generated by the full version of Kling 4.0, launching in October

    Kling officially released a short film generated by the full-featured version of Kling 4.0 with all-around reference, and said this version will launch in October. The quoted content mentions that compared with Flash, the full version offers native 30-second direct output, clearer image and sound, more precise lip-sync matching, and a 21:9 cinematic aspect ratio.

  3. 阿真Irene · 创作者专栏78

    Alibaba Open-Sources Qwen-Image-2.1: One Model Covers Image Generation, Editing and Transparent Backgrounds

    Alibaba has open-sourced the Qwen-Image-2.1 image model, with a 7B visual generation component using a 32-layer Single-Stream DiT; one model covers generation, editing and native transparent-background images.

    Why it matters: The author presents extensive hands-on cases showing Qwen-Image-2.1's performance in image generation, editing and native transparency, and provides measured speeds on consumer-grade GPUs.

Sep 29Tue
  1. LTX-2 官方发布62

    LTX-2 Releases v1.4.0, Adding Chunked Long-Video Generation and Keyframe Decoding

    LTX-2 released v1.4.0, adding --chunked to Distilled, the two-stage TI2Vid, A2Vid, Dub-It and IC-LoRA, enabling ultra-long video generation with a default 97-frame window and 25-frame carry, without needing to set --chunk-pixel-frames or --chunk-carry-frames.

    Why it matters: LTX-2 v1.4.0 lists interface changes for long-video generation, chunked decoding and HDR workflows in one place, making it easier for existing workflows to assess migration costs.

  2. 公众号:短剧自习室75

    Dream Dimension and Shengshu Technology Use Vidu S2 for AI Anchor Livestream, Generating 42,000 Yuan in Revenue in 48 Hours

    Dream Dimension and Shengshu Technology used Vidu S2 to drive AI anchor Ziying through a 48-hour livestream, drawing over 310,000 cumulative viewers and a peak of over 40,000 concurrent viewers, with total revenue of 42,000 yuan. Ziying reads out new viewers' nicknames, picks up on bullet-screen topics, and sings, dances, changes outfits and switches scenes based on gifts; the stream was restarted 24 times before reaching the 80,000 business-value target.

    Why it matters: Using a 48-hour AI host livestream as a case study, observe the real-world performance and commercial conversion of real-time video models in interactive content.

  3. Kyutai · 音视频模型20

    Kyutai Releases OVIE-ft-re10k: A Novel View Synthesis Fine-Tuned Model for RealEstate10K Scenes

    Kyutai has released OVIE-ft-re10k on Hugging Face. This is a checkpoint of OVIE fine-tuned for 50K steps on RealEstate10K multi-view data, accounting for only 2% of its 2-million-step pretraining budget. The architecture is identical to the base model, and it can be loaded directly by reusing the same code path.

  4. Lightricks · LTX 模型71

    Lightricks Releases LTX-2.5-22B IC-LoRA Alpha Gen Matting Model

    Lightricks released IC-LoRA Alpha Gen based on LTX-2.5-22B, which takes ordinary RGB video as input and generates frame-aligned grayscale alpha mattes without green screens, masks, or prompts, and can handle semi-transparent elements such as hair, fur, smoke, fire, sheer fabric, glass, and water.

    Why it matters: It generates frame-aligned alpha mattes from ordinary video without green screens or prompts, offering a new alternative for matting and compositing workflows.

Sep 28Mon
  1. 可灵 Kling68

    Kling 4.0 to Be Released in October; 4.0 Flash Now Available to Ultra Annual Subscribers

    Kling announced that Kling 4.0 will be released this October, and Kling 4.0 Flash is now available to Ultra annual subscribers.

    Why it matters: Kling 4.0 announced capability upgrades including audio-visual, reference, and multi-keyframe features, allowing creators to assess room for adjusting their production workflows accordingly.

  2. IndexTTS · 语音模型60

    IndexTeam Releases Index-Echo-S2TT-9B Speech Transcription and Translation Model

    IndexTeam released Index-Echo-S2TT-9B, which uses the Qwen3-Omni AuT audio encoder plus the Index-Translate-9B decoder to convert speech into text, accepts audio or video files, and outputs bilingual subtitles with sentence-level timestamps; the packaged inference interface supports translating Chinese into English, Japanese, or Spanish.

    Why it matters: Index-Echo-S2TT-9B has been open-sourced as a complete inference package for speech transcription plus bilingual subtitles, which creators can use to assess the replacement cost of their subtitle and translation workflows.

  3. IndexTTS · 语音模型60

    IndexTeam releases the Index-Echo-S2TT-2B speech-to-text translation model

    IndexTeam releases Index-Echo-S2TT-2B, which uses the Qwen3-Omni AuT audio encoder, an audio connector, and the Index-Translate-2B decoder to convert speech into text; it accepts audio or video files and outputs bilingual subtitles with sentence-level timestamps, and the packaged inference interface supports translating Chinese into English, Japanese, or Spanish.

    Why it matters: The Index-Echo series adds a speech-to-text translation member, with its architecture, inference parameters, and in-house evaluation comparisons made public, which can be used to judge whether to integrate it into a subtitle workflow.

  4. Lightricks · LTX 模型62

    Lightricks releases the LTX-2.5-22b IC-LoRA SDR-to-HDR model

    Lightricks releases LTX-2.5-22b-IC-LoRA-SDR-To-HDR, which converts 8-bit SDR video to 16-bit HDR and outputs EXR frames in addition to an HLG master.

    Why it matters: Lightricks turns SDR-to-HDR into an IC-LoRA based on LTX-2.5 and provides a ComfyUI workflow and EXR output path, so it can be integrated directly into post-production pipelines.

Sep 27Sun
  1. Lightricks · LTX 模型71

    Lightricks releases the LTX-2.5-22b IC-LoRA old-film restoration model

    Lightricks releases LTX-2.5-22b-IC-LoRA-Restore on Hugging Face, an In-Context LoRA based on a reference video that renders old footage — low-resolution, low-bitrate, tape, or black-and-white/sepia scans — into clean, color, high-definition footage, removing compression damage, flicker, dirt, and scratches.

    Why it matters: Lightricks open-sources an old-film restoration LoRA, providing a two-stage workflow from 480-line scans to 4K and measured gains, which can be compared with existing restoration pipelines.

Sep 25Fri
  1. AI Business42

    Google Adds TTS Models to Gemini 3.8, Strengthening Voice AI Capabilities

    On September 23, Google launched Gemini 3.8 TTS and Gemini 3.8 Flash-Lite TTS, calling them the most capable audio generation models available today. The former is aimed at creative direction and character design, capable of generating new voices from scratch using natural language prompts, covering scenarios such as games, audiobooks, and podcasts, and supporting customization of characters, accents, and voice features across more than 100 languages and dialects; the latter is used for AI dubbing, content creation, and voice agents.

  2. Google DeepMind67

    Google DeepMind Releases Gemini 3.8 Live with Live Avatar

    Google DeepMind released Gemini 3.8 Live with Live Avatar, adding near-real-time video generation to the native real-time dialogue model to achieve dynamic visual avatars with lip sync, natural expressions, and smooth turn-taking.

    Why it matters: Gemini 3.8 Live adds real-time digital human capabilities, supporting lip sync in 97 languages and background tool calls, enabling enterprises to evaluate scenarios such as virtual customer service.

Sep 23Wed
  1. Google DeepMind72

    Google Releases Gemini 3.8 Flash TTS and Flash-Lite TTS

    Google DeepMind has released two speech synthesis models: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. The former is designed for character design and line-by-line performance direction, while the latter targets high-volume, low-cost dubbing and voice agent scenarios.

    Why it matters: The two TTS models transform voice generation from fixed presets into a line-by-line directed creative workflow, enabling dubbing and audiobook teams to evaluate migration accordingly.

  2. 阿真Irene · 创作者专栏78

    Hands-on with Vidu S2: Case Breakdown of the Real-Time Interaction and Real-Time Editing Model Lines

    Vidu S2 was released on September 15, with the paper published on September 10. The web version, app, mini-program, and API are all open for use, featuring two models: real-time interaction and real-time editing. The author tested the real-time interaction model S2-Avatar, which supports voice conversation, real-time motion response, and reference image upload for interaction, and can save 720p interaction videos. Tests included virtual idol livestreaming, selling Cantonese morning tea, English conversation, and full-body scenes; after being interrupted mid-conversation, the character could pick up new topics.

    Why it matters: The author uses extensive hands-on cases to break down Vidu S2's two capability lines — real-time interaction and real-time editing — providing a basis for assessing the viability of digital human livestreaming and real-time outfit-change scenarios.

Sep 22Tue
  1. 腾讯混元 · 公开公众号 RSS72

    Tencent Hunyuan Releases Hy Image3.5 Preview, a Professional-Grade Image Generation Model

    Tencent Hunyuan releases Hunyuan Image 3.5 Preview (Hy Image3.5 preview), positioned as a high-value professional-grade image generation model. It supports text-to-image and image-to-image generation, up to 5 reference images per upload, up to 2K resolution output, and multi-turn conversational creation.

    Why it matters: Tencent Hunyuan releases a preview of its professional-grade image generation model, providing 2K output, API pricing of 0.15 yuan per image, and internal blind test comparison results, which can be used to assess cost and image generation capability.

Sep 21Mon
  1. 腾讯 ARC · 创作模型60

    Tencent ARC Releases WorldCrafter-Fast Video Generation Weights

    Tencent ARC released self-contained WorldCrafter-Fast weights on Hugging Face, supporting image-to-video, text-to-video, and an interactive image-to-video demo, with inference possible without WorldCrafter-Base files.

    Why it matters: WorldCrafter-Fast has released standalone weights, with sampling step configurations for both image-to-video and text-to-video made public, making it easier for creators to evaluate integration costs.

  2. GitHub:AI视频开源项目发现62

    Tencent Hunyuan Open-Sources Prism: Dynamic Sparse Attention for Native 2K Joint Video-Audio Generation

    Tencent Hunyuan, in collaboration with Fudan University and Zhejiang University, proposed Prism, a dynamic sparse attention framework for natively training 2K joint video-audio generation models. It organizes token sequences into spatiotemporal macro-blocks and dynamically allocates block shapes based on local content, achieving a 2.5x training speed improvement over full attention with higher generation quality.

Sep 20Sun
  1. 王若风 · 开源创作项目分享62

    Qwen Open-Sources Qwen-Image-2.1: 7B Unified Text-to-Image and Image Editing with ComfyUI Support

    The Qwen team open-sourced Qwen-Image-2.1, unifying text-to-image and image editing into a single 7B parameter checkpoint, with native support for 2K generation and RGBA transparent output with alpha channel, up to 10 reference images, and local editing via circle, doodle, or independent mask selection.

  2. 通义千问 · 图像与音频模型62

    Qwen Open-Sources Qwen-Image-2.1 and Image Editing Prompt Rewriting Model

    Qwen open-sourced Qwen-Image-2.1, a unified text-to-image and image editing model with a 7B parameter, 32-layer Single-Stream DiT for visual generation. This version supports generating standard and transparent RGBA images, up to 10 reference images, local editing via circle selection or masks, and maintains consistency for characters and products.

    Why it matters: Qwen open-sourced a unified text-to-image and image editing model, along with a prompt rewriting model, allowing creators to evaluate how to integrate editing workflows.

  3. 通义千问 · 图像与音频模型62

    Tongyi Qianwen Open-Sources Qwen-Image-2.1 Image Generation and Editing Model and Prompt Rewriting Model

    Tongyi Qianwen open-sourced Qwen-Image-2.1, a unified text-to-image and image editing model with a 7B parameter, 32-layer Single-Stream DiT for visual generation, balancing generation quality, inference efficiency, and versatility.

    Why it matters: Tongyi Qianwen open-sourced an image generation and editing model, along with a prompt rewriting model, allowing creators to evaluate image generation and editing workflows.

Sep 19Sat
  1. Pika60

    Pika releases new creation platform, Relight Media supports adjusting video lighting

    Pika has released a new AI creation platform, which the company says is aimed at creators and pursues output that meets strict standards. Using Relight Media as an example, the company explains that beyond changing the time in a video, it can also adjust the color, intensity and direction of lighting, and invites users to try it on the new Pika.