Skip to content
Date and source

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

Oct 16Fri
  1. AITNT · AI创作资讯83

    Kling 4.0 officially launched, supporting 30-second single-take shots and multi-frame references

    Kling AI officially launches the video model Kling 4.0, supporting direct output up to 30 seconds; previously, in late September, 4.0 Flash early access was opened to some Black Gold annual card and Ultra annual subscribers. The author tested scenarios including city chases, circuses, elevator fights, and figure skating, and believes it has improved in understanding physical laws, complex action transitions, and camera scheduling.

    Why it matters: The author tested 30-second direct output, complex fight scenes, and Omni Reference, and provided a cost reference of 450 credits for 1080P 30 seconds, which can be used to evaluate the production workflow.

  2. Reddit StableDiffusion · 创作者与工作流分享22

    Minimax H3 vs. Seedance 2.0 camera movement comparison test

    A user compared the multi-camera movement performance of Minimax H3 and Seedance 2.0 using the same 15-second martial arts training prompt. The prompt required presentation in the style of a 1986 American martial arts film, including continuous camera movements such as dolly-in, lateral tracking, 180-degree orbit, and low-angle tracking, while keeping the character's face, clothing, and body type consistent. The test did not give a clear conclusion on which was better.

  3. Hugging Face Daily Papers23

    Velox: Making world models more efficient with spatial memory

    Velox is a training-free attention router that uses frustum tile correspondences in streaming geometry for region-level context selection, solving the problem that KV caches in chunked video diffusion models slow down as rollouts grow. At the 48th chunk, Velox accelerates attention by 4.36× and denoising by 2.43×, and on the WBench navigation test set across all 158 cases, its average score difference from dense LingBot-Fast-Cam remains within 0.15.

  4. Hugging Face Daily Papers24

    RPO: Routing policy optimization for online diffusion reinforcement learning

    To address the high rollout optimization cost caused by perturbed reverse denoising trajectories in diffusion model reinforcement learning, RPO turns attention routing into stochastic exploration actions, samples key blocks according to native attention weights at several denoising levels, and uses a dual-rollout estimator to provide advantage signals. On text-to-image and text-to-video tasks, RPO improves generation quality while training efficiency is nearly 4× higher than Flow-GRPO-type methods, and at inference time it restores dense attention and retains the original deterministic sampler.

  5. Fstoppers60

    Is AI color grading good enough? Fstoppers tests Imagen Video

    Fstoppers author Patrick Hall tested Imagen Video's automatic color grading capability with two real projects. In a multi-camera commercial test, Imagen completed in about 35 minutes color grading that would originally take several days of manual work, and about 20% of the clips still required manual adjustment; in a three-camera podcast project, after the update it only processed materials on the timeline, reducing the time from nearly 4 hours to under 40 minutes.

  6. Hugging Face Daily Papers22

    Speculative Forcing: a speculative forcing framework for efficient autoregressive video generation

    Speculative Forcing treats intermediate clean estimates in autoregressive video generation as drafts, predicting which tokens can be accepted early and frozen as final output, thereby skipping subsequent denoising steps. On LongLive, it achieves the quality of a four-step generator with about 1.7 denoising steps, a speedup of over 1.5x, reaching 41.7 and 37.8 fps; on Lingbot-World 2.0, it maintains quality with about 2.5 steps, a 1.35x speedup, reaching 24.5 fps.

  7. Reddit StableDiffusion · 创作者与工作流分享20

    How to create a consistent AI arborist character for 15-second vertical short videos: the best ComfyUI / RunPod workflow?

    A creator asked for help on Reddit, hoping to build a reusable workflow with ComfyUI and RunPod to batch-generate roughly 15-second 9:16 vertical short videos, always featuring the same photorealistic arborist character, changing only the scene, tree species, dialogue, and camera movement.

  8. ComfyUI Wiki · 工作流与开源创作分享73

    Alibaba TaoLive AIGC releases the TaoMate-H3-FL2VA first-and-last-frame control LoRA

    The Alibaba TaoLive AIGC team released TaoMate-H3-FL2VA on October 15, a 3-step LoRA that adds first-frame and last-frame control to MiniMax H3's streaming TaoMate-H3 runtime; the team calls it the first known real-time streaming first-and-last-frame-to-audio-video (FL2VA) model.

    Why it matters: TaoMate-H3-FL2VA integrates first-and-last-frame control into streaming audio-video generation, and a ComfyUI conversion version is already available to run locally, making it easy for creators to evaluate whether to incorporate it into their existing production pipelines.

  9. Wireflow · AI短剧与视频制作教程48

    How Claude generates video via MCP: a pipeline from script to finished film

    Claude itself cannot generate video, but it can call video models via MCP and return a finished MP4. In a four-node pipeline built on Wireflow, Nano Banana Lite generates keyframes, Gemini Omni Flash 1.1 and Kling O3 each produce 5-second clips, and Video Editor composites a 1920x1080 finished film. The whole process takes 101 seconds and consumes 176 credits.

  10. 虎嗅 · AI影视与短剧产业74

    Hengdian joins hands with Volcano Engine with 235 million yuan to win the bid for an AI film and television service platform, aiming to build an online AI film and television city

    A notice from the Dongyang branch of the Jinhua City Public Resource Trading Center in Zhejiang shows that a consortium of Dongyang Deep Space Engine (a ByteDance subsidiary) and Volcano Engine won the bid for the Hengdian Film and Television AI Service Platform construction project with a total of 235 million yuan; the tender was launched in July. The project centers on the Hengdian Film and Television large model and will build 7 vertical-scenario models, 20 specialized-scenario models, 17 film and television industry agents, 9 categories of high-quality datasets, and a consumer-facing Hengdian Film and Television App. The tender documents set requirements including digital extras generation, actor style transfer, and AI-assisted dubbing.

    Same story, the pick is 《Hengdian joins hands with Volcano Engine on a 235 million yuan AI film and television service platform, while the number of offline film crews drops sharply over the same period》

  11. VibeClips · AI视频创作教程20

    Privacy and security of AI video generators: how VIBE handles your photos, prompts and videos

    VIBE's privacy policy stipulates that uploaded photos are only sent to the AI model APIs needed to generate videos (such as Kling, Veo, Sora, Seedance, WAN), are not used to train models, are not used for facial recognition, are automatically deleted after being stored for a maximum of 90 days, and users can also manually delete them at any time. Prompts and generated videos are retained in the account until the user deletes them, and personal data is not sold; the terms of service grant users a license to use generated videos for personal and commercial purposes, and users retain ownership of the original prompts and images.

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

    Writing the script with Claude and drawing backgrounds with Gemini: how a 32-second AI video became a 70-second Halloween piece

    The author gave the pixel animation script to Claude and the backgrounds to Gemini, expanding an originally 32-second video into a 70-second Halloween short film "カボチャのみちしるべ". After telling Claude directly "I don't think it's impressive yet", it pointed out four missing points: "event moments", character performance, composition changes and musical ups and downs, and based on this added three highlights: 22 pumpkins lighting up in sequence, a swarm of 120 bats, and a window-shadow ending.

  13. 娱乐资本论 · 澎湃商业与影视趋势70

    娱乐资本论's comparative review of 10 AI video canvas platforms: functions diverge, with different emphases for three user groups.

    娱乐资本论 releases the functional review of AI video canvas, testing 10 platforms (AniShort, 巨日禄, LibTV, 纳逗Pro, OiiOii, Seko, ShotLab, TapNow, 万兴剧厂, 小云雀), and organizes functional differences according to three user categories: serialized comic dramas, professional film/TV, and cross-border e-commerce.

    Why it matters: The review covers 10 AI video canvas platforms, sorting out functional differences for three user categories—comic dramas, film/TV, and e-commerce—to help creators choose according to their needs.

  14. TheoreticallyMedia · YouTube 创作视频62

    ByteDance world model remakes "King of Comedy"; OmniHuman 2.0 and Krea 3 debut at the same time

    ByteDance's unreleased world model demonstrated more than ten new camera angles of "King of Comedy" scenes generated from a single hand-drawn camera path at a Douyin gala, with authorization from Stephen Chow. In the same period, OmniHuman 2.0 launched as an API through BytePlus, available only outside the United States, with a maximum of 5 minutes per clip, 720p, and a price about 40% lower.

  15. Higgsfield 官方博客38

    7 best Veo alternatives: how to maintain consistency during generation

    The official Higgsfield blog compared 7 Veo alternative tools, focusing on the consistency of characters, clothing, and scene details across multiple shots. Cinema Studio 4.0 supports up to 50 reference assets, 30-second clips, and 1080p output, with a 10-second 720p clip including audio consuming 70 credits; Seedance 2.5 supports 30-second native stereo and 50 references, and Kling 3.0 can generate up to 6 shots at a time.

  16. かみもと · 本地AI视频工作流62

    23 MiniMax H3 custom nodes selected: a summary of long-video, editing, audio, and asset features

    The author organized 23 MiniMax H3 custom nodes into six categories: long video and continuous generation, performance and editing, image quality correction, audio and dialogue, reference assets and character motion control, and assets and production support, with links to each repository's README and workflows.

  17. Reddit StableDiffusion · 创作者与工作流分享62

    A workflow for drawing stick figures with Valtz on Mac and generating AI videos

    Valtz offers a workflow on Mac for generating AI videos from stick-figure animation: draw a simple stick figure on the canvas, insert Position and Zoom keyframes in the Crop and Position panel, use the joystick to move the Markup Layer and record displacement, then select Linear Interpolation on the first keyframe in the timeline, and Valtz automatically interpolates frames to generate smooth animation.

  18. Hugging Face Daily Papers27

    Execute first, remember later: diagnosing memory failure in video world models and the MIST benchmark

    The research proposes the MIST benchmark, which separates "whether the scene is set" from "whether instruction changes are executed," finding that the bottleneck of video world models lies in unexecuted instruction changes rather than forgetting. The accompanying method M3-Agent recursively rewrites input instruction details, raising the MIST reveal score by more than three times, improving the T score on WRBench by 26.1% and the U score on MBench-T by 52.7%.

  19. note · aicreatorpathnet · 图像视频创作分享62

    Using MiniMax H3 to generate a full 125-shot, roughly 20-minute AI historical film

    The author used MiniMax H3 to create the roughly 20-minute, 125-shot historical film 《本能寺から山崎へ ― 信長・光秀・秀吉の十一日》, covering the Honnō-ji Incident through the Battle of Yamazaki. It was released on YouTube on October 12, 2026, in a Japanese version (19 min 38 sec) and an English version (19 min 29 sec), both in 4K.

  20. Curious Refuge film gallery42

    DEFCON 1: the AI short 《DEFCON 1》 depicts a nuclear crisis countdown using Seedance 2.5

    The AI short 《DEFCON 1》, created by AI VIDEOS, fictionalizes a nuclear emergency as two warheads approach Seattle and New York. With no narration throughout, it follows military crews, the president, and fleeing civilians minute by minute through a failed interception and the final retaliatory order. The film was generated using Seedance 2.5 on Higgsfield and edited in DaVinci Resolve.

  21. Curious Refuge film gallery22

    《Low Orbit》 AI sci-fi short: a night train heading toward an abnormally close moon

    The AI surreal sci-fi short 《Low Orbit》, created by Frankly Lying, turns an ordinary night train commute into an unsettling encounter with an abnormally close moon. The soothing instruction "Everything is fine. Do not slow down" recurs throughout, creating dreamlike tension between routine daily life and approaching danger. The film is a screening selection of the Higgsfield Global Film Festival.

  22. 我爱计算机视觉 · 公开公众号 RSS31

    HKUST × Tencent Video propose Adaptive Reward Routing: audio-video diffusion model post-training no longer fixes reward routing in advance

    MMLab@HKUST of HKUST, Tencent Video, and the University of Hong Kong propose Adaptive Reward Routing, which, on the DiffusionNFT framework, recomputes multi-reward routing for audio-video diffusion models in real time during training rather than fixing it once from the base model.

  23. Reddit StableDiffusion · 创作者与工作流分享65

    Author trains H3-Anti-Burn LoRA to mitigate over-sharpening and burning after chaining H3 clips

    The author releases H3-Anti-Burn LoRA to mitigate the problem of footage becoming progressively over-sharpened and burned when H3 clips are chained into longer videos. The LoRA is loaded only on continuation clips, with an initial weight of 0.5. In a chained test of about 37 seconds, facial texture change for lightx2v 8-step dropped from +44% to 0%, HyperFlow from +25% to −1%, and Larry Turbo from +261% to −5%.

    Why it matters: To address over-sharpening and burning after chaining H3 clips, the author provides a loadable LoRA with specific weight recommendations, along with multiple sets of before-and-after comparison data.

  24. Reddit StableDiffusion · 创作者与工作流分享62

    Workflow for character replacement with Minimax in ComfyUI

    The author reproduces the common TikTok character-swap videos in ComfyUI using Minimax, replacing real-person dances with animals or other characters. The approach is to load MiniMax-H3-Character-Swap-LoRA on the default ref2va template workflow and first use the default Qwen 2.1 edit template to generate a single-frame guidance result.

  25. Reddit comfyui · 创作与工作流分享60

    Ome Omy CUT's former Filmmaker version is now free; Fast H3 Omni integration coming soon

    Ome Omy Productions has made the CUT version, previously part of the $5 Filmmaker tier, free. The author describes it as a local MiniMax H3 / ComfyUI timeline editor that can generate shots, preserve different takes, edit and trim footage, extend scenes, patch discontinuities, and arrange results on the timeline. It also supports saving projects, comparing takes, and exporting finished videos.

  26. Reddit comfyui · 创作与工作流分享38

    Template site prompts perform poorly when migrated across models: LTX-2.5 requires rewriting in the model's own style rather than direct copying

    Copying prompts from template sites to test LTX-2.5 consistently yielded poor results until the user compared them with the reference prompts built into the official workflow. The built-in prompts are approximately 200 words and include explicit camera movements, pacing nodes, and on-screen text typography details (e.g., 'pure white sans-serif, wide spacing, no glow, no shadow'), whereas most shared prompts are tag-list style written for Midjourney/Ideogram. Rewriting the same scene in the model's own expected style produced dramatically different results.

  27. Monster社区 · 中文ComfyUI实测与工作流62

    Image Tagger 1.3.0 released: video prompt reverse-engineering, translation, and batch export

    Image Tagger V1.3.0 chains prompt reverse-engineering from reference videos, bilingual editing, batch export, and long-video segmentation into a single workflow, with this release focusing on prompt configuration for target models such as Seedance. To use it, first create a project and import online or local videos, then select the model series and conditional modes (T2V, I2V, full-modal, video extension, video editing) based on the intended generation use case. After reverse-engineering, prompts can be translated between English and Chinese for side-by-side editing and saving.

  28. What’s New in Generative Media22

    Generative media today: AI cuts an $80 million film budget down to $1.5 million; Amazon in talks to acquire Decart for $7 billion

    The director of *Christmas Witch Merry* estimated the film's traditional production cost at approximately $80 million, but it was completed for just $1.5 million with AI assistance. Amazon is in talks to acquire AI startup Decart for approximately $7 billion. MiniMax-H3-Turbo released a fixed-step checkpoint supporting text-to-video with audio and first-frame-to-video generation.

Oct 15Thu
  1. 可灵 Kling66

    Kling 4.0 Fully Opened, Adding MCP and CLI Integration

    Kling has announced that KLING 4.0 is now fully open. Users can access it through the updated creation interface on kling.ai, or integrate it into video creation workflows via MCP and CLI. The company also previewed upcoming 10-bit HDR video output at 4K and 1080p resolutions, as well as API access for enterprise clients and developers.

    Why it matters: Kling 4.0 is now fully open with new MCP and CLI integration options, allowing creators to evaluate integration choices for their existing production workflows.