Skip to content
Date and source

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

Oct 13Tue
  1. Loomance · 本地AI动画实测60

    Wan 2.2 first/last frame test: three runs with unrelated images yield three transitions

    The author ran an experiment locally on an RTX 5070 12GB using Wan 2.2's First → Last Frame mode, with a hand-drawn spotted camel-like creature as the first frame and a photorealistic apple as the last frame; the same prompt run three times produced three different transitions: placing the drawn character into the castle scene of the last frame, staying in the drawn world before a quick cut, and a slow dissolve between the two worlds.

  2. ComfyUI Wiki · 工作流与开源创作分享62

    LightX2V team releases LightVAE: a compressed video VAE decoder with ComfyUI nodes

    The LightX2V team introduced LightVAE, which compresses pretrained video VAE decoders using temporal low-rank convolution and block pruning while retaining the original latent interface, so the original encoder can still be used. It covers Wan2.1, Wan2.2 and MiniMax-H3, offers six checkpoints on Hugging Face, and comes with ComfyUI nodes, six generation workflows and decoding benchmarks.

  3. Curious Refuge film gallery22

    41 Frames' AI horror short "John O' Lantern": a pumpkin family harvests human heads for the holiday

    The AI horror short "John O' Lantern" by 41 Frames builds a world where a pumpkin family harvests human heads to make pie, brew soup and celebrate Halloween. The film contrasts its bloody premise with cheerful family dialogue and familiar holiday rituals, subverting Halloween traditions through dark humor. The film is an entry in the Curious Refuge AI Horror Film Competition, and its ending delivers a twist as the pumpkin family tells a horror story about humans cooking pumpkins and decorating houses with pumpkin children.

  4. Curious Refuge film gallery22

    Muzappar Osman's AI horror short "Don't Cut His Nails"

    Muzappar Osman released the AI horror short "Don't Cut His Nails," about Mira, whose obsessive-compulsive pursuit of perfection leads her to cut her new partner's nails, after which a seemingly perfect romance slides into a nightmare. The film opens with confessional narration and light dating conversation, then turns to an unsettling supernatural truth, transforming an intimate morning after into a dark story about obsession and crossing boundaries.

  5. Hugging Face Daily Papers12

    Scaling online optimal transport minibatches to improve flow matching training stability

    Researchers propose a scalable approximation of online entropic optimal transport that substantially increases the pairing batch size between source samples and data in flow matching training while retaining the benefits of optimal transport. They theoretically prove that the method strictly reduces full stochastic gradient variance and integration error in both the unregularized and entropy-regularized cases, and that the improvement grows as the online coupling batch size increases. Experiments on single-cell gene interpolation flows and image flow training verify that larger online coupling consistently improves training stability and final generation performance.

Oct 12Mon
  1. 娱乐资本论 · 搜狐公开专栏38

    AI Short Dramas Are Not a "Hunt" but the Daily Life of 800 Million People

    In response to trending topics such as "AI short dramas are hunting middle-aged men," the article points out that AI short dramas are already the daily life of 800 million people: as of May 2026, micro-drama users nationwide reached 851 million, 66.7% of internet users, and in the first eight months of 2026, 430,000 micro-dramas went online, with AI dramas accounting for over 90%. The widely circulated comparison that "Chinese AI short dramas account for 55% of tokens and US programming 58%" is said to be a mismatched methodology lacking totals and authoritative sources.

  2. 短剧自习室 · 文章60

    How the "Not Heartburn" Meme with 2.09 Billion Views Became a Byproduct of the AI Short-Drama Assembly Line

    As of October 12, the topic "Not Heartburn" had 2.09 billion views on short-video platforms; its source is a batch of AI street-food short dramas with highly similar plots and lines, repeatedly outputting lines such as "It's a bit pricey, but eating it doesn't give you heartburn" through the fixed template of "a conscientious small shop fighting unscrupulous competitors."

  3. 腾讯 ARC · 创作模型60

    Tencent ARC Open-Sources Multi-Person Video Generation Model Multics

    Tencent ARC released Multics on Hugging Face, achieving scalable multi-person video generation through layered communication. The repository includes the complete 4-player Multics DiT checkpoint, configuration, Wan2.2 VAE, UMT5 text encoder, and tokenizer; the same checkpoint supports 6-, 8-, and 10-person generation without retraining. The weights are distributed under the Apache 2.0 license.

    Why it matters: Tencent ARC open-sources the multi-person video generation model Multics, where the same weights can scale to 10 people, letting creators evaluate generation options for multi-person scenes.

  4. What’s New in Generative Media17

    Generative Media Weekend Roundup: Iranian AI Fake News, fal Integrated into ChatGPT, New MiniMax H3 Tools

    fal's minimax/h3-max/relight can now be called directly in ChatGPT and Codex, letting users generate media without leaving the conversation. ComfyUI-MiniMaxH3-Studio supports turning descriptions into MiniMax H3 videos within ComfyUI, with a built-in editor for editing and rendering. In addition, PlateExtract, fine-tuned on Qwen Image 2.1, can extract transparent PNG foregrounds from composite images.

  5. TelosCompass · note AI视频创作实践34

    Making Product AI Videos with DomoAI: Write the Surrounding Motion First, Then Move the Product

    When generating product videos from photos with DomoAI, you should first write the motion of surrounding elements such as light, shadows, and steam, then decide whether the product itself moves. DomoAI's official blog suggests that products with dense text labels are best kept completely still, and if rotation is needed, keep it within 15 degrees and specify the stopping position; for scenes with people, keep only one action and clearly state the speed and pose.

  6. note · abcxy · 图像视频创作分享50

    For Complete Beginners: A Tutorial on Generating Video from a Single Image with MiniMax H3

    MiniMax H3 is a video generation AI model released by MiniMax in July 2026; it can generate video and audio simultaneously (including Japanese dialogue) and supports Turbo LoRA to reduce generation steps. In ComfyUI 0.30.0 and above, using the official template, you can generate a short video from a single image by simply placing the image (first frame) and describing the motion with a prompt, making it suitable for complete beginners.

  7. note · hatano_yusuke · 图像视频创作分享22

    3 shortcuts for automatically generating video with generative AI

    There are three shortcuts for building video generation applications with generative AI: directly integrating commercial video generation APIs such as Sora, Runway, Luma, and Pika; self-hosting open-source models such as Stable Video Diffusion with cloud GPUs; or using iPaaS and no-code tools to string together an automated pipeline of text generation, video generation, and speech synthesis APIs. Each of the three involves trade-offs in development speed, cost, and customizability. You can first use APIs or no-code tools to quickly build a prototype, then migrate to a self-hosted pipeline as needed.

  8. note · wise_azalea8300 · 图像视频创作分享78

    Comparison of 44 ComfyUI video upscaling and restoration methods, with an all-in-one comparison workflow

    The author compared 44 video upscaling and restoration methods on ComfyUI, all running locally for free, and distinguished between two scenarios: restoration and upscaling. Restoration targets low-quality footage from small cameras, surveillance, and the like, while upscaling targets upgrading HD to FHD, 4K, and 8K and recovering skin texture and eyelash detail. The article releases a large-scale comparison workflow that can run all 44 methods in one go, making it easy to pick suitable methods for different content such as people, landscapes, or animation.

    Why it matters: The author separates the two needs of restoration and upscaling, uses the same footage to compare 44 methods side by side, and releases a ComfyUI workflow that can run in one click.

  9. note · comfylab · 图像视频创作分享75

    Hands-on test of MiniMax H3's RefMod node: keep the same character appearing repeatedly without training a LoRA

    The author tested MiniMax H3's RefMod custom node in ComfyUI. It compresses character images, motion videos, or audio into reusable reference files, requires no LoRA training, is used much like a LoRA loader, allows adjusting strength, and can combine multiple references to put characters in the same frame.

    Why it matters: The author tested MiniMax H3's RefMod custom node and provides the complete parameters and prompt-writing approach, from installation and building a reference library to integrating with the official T2V workflow.

  10. 可灵 Kling42

    How Kling 4.0 creates viral AI music videos

    How do you make an AI-generated music video that gets people talking? 👀🎬 Simon Meyer, Creative Director and co-founder of PROMPTR Studios, got early access to Kling 4.0 to test its new capabilities. The music video he created went viral, and viewers wanted to know how he did it. In this episode of Creator Off Script, Simon breaks down his creative process, shares how he made the AI visuals look real, and reveals the Kling 4.0 features that impressed him most.

  11. note · hatano_yusuke · 图像视频创作分享20

    Hyperframes video generation pitfall guide: 5 key points

    When developing an AI video generation application with Hyperframes, you can reduce pitfalls in five areas: validate early with low resolution and short frame counts, then scale up gradually; fix seed values and use control networks to maintain consistency between frames; keep prompts concise and set negative prompts; use asynchronous processing, queues, and retries to handle API timeouts and rate limits; and pin model and dependency versions and verify updates in a test environment.

  12. note · major_myrtle6198 · 图像视频创作分享22

    Running MiniMax H3/h3.c at maximum quality with no quantization on an M1 Max 64GB

    On an M1 Max (64GB) Mac, the author ran h3.c, the C port of MiniMax H3, in BF16 with no quantization, successfully generating a 10-second video without triggering out-of-memory (swap), but it took 43.2 hours. The full model is about 134GB and requires about 150GB of available storage, reading and computing in chunks from the SSD via an SSD streaming mechanism.

  13. note · 7god_ebisu8g2sdr · 图像视频创作分享22

    Even PC beginners can use ChatGPT to make AI videos while chatting: from PixVerse prompts to a YouTube Shorts hands-on

    A PC beginner used ChatGPT to generate and revise PixVerse prompts to make a 5-second vertical ASMR video of a glass pumpkin cutter. The prompt ChatGPT initially gave was too long and did not work well in PixVerse, so the author directly gave the feedback "the prompt is too long," and ChatGPT promptly produced a short version containing key points such as "transparent orange glass pumpkin on a wooden cutting board, macro close-up."

  14. note · takah_note · 图像视频创作分享8

    Online Course | Real AI Video Production That Includes Failure

    A creator shares real experience making online course videos with AI: dubbing with a voice clone of their own voice, but after recording 4 versions with emotional ups and downs, no difference could be heard, so they ultimately kept the original voice and instead created a lively feel through writing techniques such as ending with questions.

  15. note · koza_kura3 · 图像视频创作分享20

    Can't Make AI Animation Without a High-End Computer? The New Common Sense That Ordinary PCs Can Also Get Started

    Making AI animation does not necessarily require a high-performance computer costing 200,000 to 300,000 yen. With cloud-based image generation AI and video generation AI, heavy computation is handled on the provider's servers, so an ordinary laptop or phone can also give it a try. It is recommended to first use browser-available tools, start with short 3-5 second animations, and confirm free quotas and commercial-use terms before deciding whether to buy a machine.

  16. 量子位 · 搜狐公开专栏38

    ByteDance World Model Debuts: Generate a Freely Explorable 3D World From One Sentence

    ByteDance showcased its world model for the first time on the tech segment of Douyin Group's "2026 Wonderful Night," where inputting one sentence generates a 3D world that can be freely flown through, interacted with in real time, edited at the object level, and connected across scenes. The model has not yet been officially released; the event also demonstrated training and validation of autonomous driving and household robots in AI-generated virtual environments.

  17. cnBeta22

    Musk on Robot Demonstrations: Pre-Choreographed Movements Don't Count as Real Skill; Optimus Still Needs to Overcome Challenges Like Screwing in Screws

    In an X Takeover interview, Musk said many robot demonstration movements are pre-choreographed, and that a robot only has real skill if it can handle situations it has never seen; his own Optimus still needs to overcome challenges in strength, precision, and touch for tasks like screwing in screws. He revealed that the next Starship test flight will hopefully attempt to catch the upper stage, enabling recovery of both the booster and the ship, and said AI music videos can already rank in the top ten or twenty he has seen, with some using Grok 4.5.

  18. SuSu_酥酥38

    Media Storm's Tim Explains ByteDance's World Model in Depth

    Media Storm founder Tim demonstrated ByteDance's latest world model in a talk: one sentence can generate a freely explorable 3D scene, with physical effects such as characters walking and liquid splashing responding in real time, and further instructions can be given to modify it. The model supports turning old movies into "living worlds" with freely controllable camera movement, seamlessly traveling through different scenes via portals, and can be used for simulation training of robots and autonomous driving; combined with VR glasses, it can also provide high-fidelity immersive simulation for healthcare, training, and more.

  19. note · mintjuku · 图像视频创作分享15

    Mint Juku uses Kling AI to have children create AI videos; YouTube channel @mintjuku goes live

    Children at Mint Juku created AI videos using Kling AI, and the works have been released on the YouTube channel @mintjuku. During production, the children practiced putting their own ideas into words, and learned to face failure and to set an "80 points is enough" standard. Released works include "Thank You for Flying Angel Airlines," "Madame Gunma's Monologue," and "Self-Love Is Important," among others.

  20. 文娱价值官 · 澎湃商业与影视趋势43

    "Inertia": The historical logic and practical power of AI short dramas.

    Hongguo Short Drama's DAU reached 168 million, a year-over-year increase of 107%. AI short dramas have become the most watched and interacted content on its platform. Traditional video platforms continue to lose users. The combined daily active users of Tencent Video, iQIYI, Youku, and Mango TV are only 155 million, with MTS declining year-over-year by 31%, 27%, 47%, and 24% respectively. The article argues that AI short dramas are the most fitting result of the "inertia" logic, and the convergence of short video consumption habits and AI technology drives their industrialized production.

  21. 影视独舌 · 搜狐公开专栏32

    AI short dramas diverge into three paths; Hongguo wins the market by "not striving to advance"

    "The Honest Son-in-Law of the Tang Dynasty: My Sister Is the Crown Princess Season 3" set a new all-time high popularity record for Hongguo AI simulated-human short dramas, ending the consecutive reign of "Everyone's Been Reborn, Who Still Pretends to Be a Rich Second Generation?" The drama reinterprets court intrigue through the logic of household scheming; AI is used only to cut costs and boost efficiency, while camerawork and scenes still follow the mature conventions of live-action short dramas. Hongguo, Douyin, and traditional long-video platforms have thus diverged into three AI short drama paths, and matching platform needs matters more than blind innovation.

  22. Simon聊AI32

    Making a short film with Opus 5.5 plus MiniMax H3 costs about 10 yuan

    Making a short film with Opus 5.5 plus MiniMax H3 costs about 10 yuan, and you can decide the topic, style, and script yourself. But even with faster production, a poorly chosen topic still gets no viewers; the author now spends far more time on topic selection than on production. He believes self-media ultimately competes on topic selection, and so far he has not seen a model with a deep understanding of topic selection.