Researchers proposed a training-free sparse allocation method, ALLOT, which treats the combination of each layer and each denoising step as a unit, estimates its contribution to the final video error through unit probing and trajectory calibration, and jointly optimizes the density of each unit under a measured latency budget.
Aizawa posted five scene illustrations on X from the morning of October 21 to late night on the 22nd, and freely distributed corresponding ChatGPT image generation prompts, covering opening curtains in the early morning, reading a letter during the day, napping on the sofa in the evening, wrapping in a blanket at night, and clipping a moon-shaped bookmark late at night. The prompts require attaching a reference image of an adult character aged 21 or older that the user has the right to use, to maintain facial features, hairstyle, and eye color, and uniformly specify a delicate and clean anime style and horizontal composition.
A creator shared the real experience of continuously producing AI videos: spending tens of thousands of yen per month on AI tools, taking 6–8 hours to produce a single video, and paid generation often produces failures such as mismatched actions, broken characters, and background changes, with failed generations also billed. Beautiful visuals are no longer a differentiator, and YouTube rules and judgment may also affect monetization.
The AI micro-drama "The Never-Disconnected Communications Squad," originally produced by China Broadcasting Network, premiered on national cable TV platforms and the "China Broadcasting Network" App, commemorating the 90th anniversary of the victory of the Chinese Workers' and Peasants' Red Army's Long March. The series focuses on June 1935 when the Red Army crossed Jiajin Mountain, telling the story of modern communications engineer Lin Yue, who travels back in time to join the Red Army communications squad and protect the radio lifeline. The work uses AI to complete scene restoration and image rendering, compressing set construction and post-production cycles and reducing production costs.
At the 2026 Taobao AIGC Content E-commerce Ecosystem Closed-door Meeting, Taobao, together with CNR, iQIYI, Migu, and Fengmang, released the E-commerce Short Drama Co-creation Alliance · Ten-Billion Quality Plan, investing ten-billion-level ecosystem support resources. Taobao Short Drama also opened an institutional onboarding entry and announced Double 11 recruitment and incentive policies, among which the October AI short drama million-yuan incentive plan offers a maximum reward of 300,000 yuan per title in revenue sharing, with a guaranteed minimum of 5,000 yuan per title for hit dramas; for live-action short drama quality exclusive cooperation, the maximum production guarantee per title is 1.5 million yuan, with a maximum playback incentive share of 1.5 million yuan.
Why it matters: Taobao shifts its Double 11 traffic budget to short drama content assets and provides specific incentive amounts for AI short dramas and live-action short dramas, which content parties can use to evaluate channel choices.
Short drama revenge narratives can be subdivided into four types: hatred of being bullied, hatred of poverty, hatred of betrayal, and hatred of being looked down upon. However, over-reliance on hatred is causing audiences to experience "hatred fatigue." The article suggests that creators should make revenge a story worth telling again, shifting toward more diverse emotional drivers such as love, responsibility, and dignity.
RubatoCodec is a variable-frame-rate speech codec based on X-Codec-2, achieving adaptive tokenization through utterance-level event budgets and streaming decoding. Its Adaptive Firing mechanism uses codec features and waveform spectral changes to divide variable-length events without needing a learned boundary predictor. At low event rates, reconstruction fidelity and intelligibility outperform fixed-frame-rate coding, intelligibility remains best on English and Japanese benchmarks, and in matched TTS comparisons WER is lower while UTMOS and UTMOSv2 are higher.
The research proposes Future Chunk Prediction (FCP), an auxiliary objective jointly optimized with teacher-forcing denoising, predicting temporally separated video chunks without seeing intermediate visual context.
The V2 INT8 checkpoint of the community fine-tuned model MiniMax H3 Singularity was uploaded on October 21 to the WarmBloodAban/Minimax-h3_Singularity repository, with the file Minimax-h3_Singularity_ref2va_v2_int8.safetensors, size 31.7 GB.
Stable Diffusion Art released a tutorial on how to use MiniMax H3's Add Guide node in ComfyUI to lock multiple keyframe images, video clips, and audio to specific frame positions on a 24fps timeline, generating a video with stereo dialogue, sound effects, and music in one pass.
Why it matters: The tutorial provides specific node settings for multi-keyframe locking and audio-visual continuation, and fixes the timestamp truncation issue in the official template.
German candy brand Ahoj-Brause, in collaboration with Berlin agency Ozmoze, launched a seven-episode vertical Halloween micro-drama, available on Instagram, TikTok, and YouTube from October 17 to 30, with episodes released several days apart. The series parodies Scream as well as Obsession and Backrooms, released in 2026, with each episode ending on a cliffhanger, and an outdoor mural in Hamburg is also incorporated into the plot.
French production company Barjac Production announced its entry into vertical short dramas, with its parent company Studio TF1 announcing the news on social media on October 20. Two of the company's vertical short dramas are competing in Vertival—which Satellifacts calls France's first vertical fiction competition—one of which has a title beginning with "Trop." The public content did not disclose the platform, budget, or release date.
Creator kingroka released the LoRA "UDA" for use with Moss Image to generate anime and digital paintings, to alleviate the strange artifacts and errors that the base model produces in digital art. The LoRA is based on about 250 images, trained for 10,000 steps, with the dataset using structured JSON bbox prompt types, and incorporates merged conversion data from three Ideogram LoRAs: anime/digital painting, Ghibli style, and retro anime.
DramaClip v0.2.2 is released, adding three toolbox capabilities: video copy extraction, speech-to-text, and text-to-speech, and includes a built-in HTTP gateway with local loopback only plus token authentication for external Agent calls.
At LA Tech Week 2026, Kling joined Oxen AI and Obsidian Studio to discuss how AI video integrates into enterprise content production—from shot creation to final delivery.
Watch the highlights to learn how Kling 4.0 supports complex production needs.
Xiaozhi Blog released a tutorial introducing how to do text-to-video with MiniMax H3 in ComfyUI, covering environment preparation, hardware and software requirements, and obtaining MiniMax H3 model files from ModelScope or Hugging Face, and provides steps for installing and upgrading huggingface-related dependencies.
Longshot Studio has been updated to 0.15.0, with multiple rounds of iteration on lip-sync and facial refinement; the API is free and runs on the user's own ComfyUI, and the code is open-sourced on GitHub. The author also released a music video, "Falling Fast," produced entirely with the studio as a demonstration.
A user built a comparison page testing the image editing capabilities of Fibo 2 edit, Klein 9b, and Qwen 2.1 at 1 megapixel with the same seed and int8 single-pass generation, concluding that Qwen 2.1 remains the best for their use case, and Fibo could be a good choice if it gains community support.
Apple's first foldable iPhone Duo will go on sale on October 23, and multiple screenshots of built-in apps have appeared on the App Store, previewing the experience of the dual-screen device when unfolded in advance; Alipay version 12.12.38 has also been adapted.
Luma's Ray 3.2 and Seedance 2.5 focus on frame-level control and synchronized audio generation, respectively, allowing users to choose based on project needs. Ray 3.2 supports HDR output, EXR export, and up to 16 keyframes, making it suitable for professional post-production; Seedance 2.5 can generate 30-second clips with synchronized sound effects, music, and lip sync in one go. Both are available through Luma and can be switched within the same workspace.
Gemini Omni Flash excels at conversational editing and multimodal input. Wan 3.0 can generate clips up to 30 seconds long and supports document input. Both have their own advantages. Luma AI's Ray 3.2 offers multi-keyframe sequences, HDR output, and EXR export. Layers (powered by Uni-1) supports single-element modification, making it suitable for production workflows. In 2025, the AI video generation market size is approximately $700 million to $2 billion, with North America leading.
The comparison of Veo 3.1 and Runway Gen-4.5 shows that when selecting AI video tools, creative teams need to consider both generation quality and post-production workflow, and Luma Ray 3.2 joins the competition with multi-keyframe control (up to 16).
In the 2026 AI video generation comparison, Seedance 2.5 and Kling 3.0 each have their strengths: the former focuses on 30-second native clips, multi-reference input (up to 50), and one-shot native audio-video generation, while the latter focuses on up to 4K output and realistic physical motion.
Douyin Group's Short Drama Copyright Center released a call for collaborations on interactive dramas and interactive film games, opening up joint distribution, joint investment, and joint IP development, seeking interactive works produced by live-action and AI. During the same period, three interactive story apps targeting overseas markets appeared, among which DramaRemix focuses on plot choices and AI video creation, Boss Reels combines short dramas with mini-games, and inStory encourages users to use AI to turn personal experiences into video stories.
Why it matters: Douyin opened a call for interactive drama collaborations, while three overseas interactive story apps launched at the same time, providing a basis for observing platform moves and product formats in the interactive short drama sector.
HeyGen Voice Turbo integrates into ComfyUI as a partner node, supporting voice cloning from a single recording and converting any script into speech. Turbo costs less than 1 cent per 1,000 characters, targeting scaled voiceover scenarios such as localized ads, training videos, product explanations, and AI digital human pipelines.
Open-source canvas Agent released v1.6.9, unifying the schema, Policy, visibility, execution binding, and approval preview of 29 tools, and removing old tool copies and duplicate context pipelines. Canvas Agent fixes focus loss when more than 8 nodes are multi-selected, and step cards add step-by-step undo; submitted generation tasks cannot be undone. The backend database Schema is upgraded to 58, adding Agent memory compression proposals and revision records; upgrading requires updating both Web and Backend.
A user discovered on macOS that the Gaia 2 (Gaia Animation) model weights at the path `/Applications/Topaz Video.app/Contents/Resources/models/coreMLCache/ganim-v1-fp16-576x672-2x-OS15-mlp` in the Topaz Video app are unencrypted.
User Sixhaunt trained a rank-16 LoRA (6,000 steps) with AI toolkit and compared two reference-image input methods for Moss Image Edit: the direct reference-image method tends to mix inputs together without a LoRA, while Ostris's editing method approximates image overlay without a LoRA and gives better editing results with a LoRA. The Ostris method renders about 45s per image, the simple reference method about 15s, and training takes about 3 times as long.
Kling 4.0 is now available on Higgsfield AI, with single-generation duration extended from Kling 3.0's 3 to 15 seconds to 3 to 30 seconds, up to 1080p with dual-channel stereo.
Why it matters: Compared with Kling 3.0, it lists upgrades in duration, keyframes, reference materials, and audio, and gives credit costs for each resolution to help assess integration costs.
YouTube commissioned NRG to survey nearly 700 users aged 14-44 who uploaded videos in the past 12 months; 72% of surveyed uploaders said they use AI on videos or images. More than two-thirds of AI video publishers said they themselves appear in the content, and 61% of respondents are interested in using AI to insert themselves into video footage. The report also mentions creators using generative AI to build complete film worlds, forming a group of "AI authors."
MattVidPro used Claude Opus 5.5 to rebuild fictional game footage generated by Seedance into a playable browser shader in 45 minutes, with the help of gokaygokay's video-to-shader skill.
MiniMax H3 Novel Pipeline released an update, adding editable reference links and a review process, and supporting recovery of extraction from saved cinematic chapter JSON. The H3 generation stage adds scene duration guidance, configurable maximum shot count, and optional shot direction, and can use Qwen to correct invalid shot post-processing notes based on validation feedback.
The author demonstrates completing a real generation in local ComfyUI without relying on a local inference GPU and without downloading models, by installing the free Forma custom node and connecting an account, then opening the image-to-video template. The solution is pay-as-you-go with no subscription, top-ups starting at $5, with 39 image and video models, 8 video workflow types, and 2 image workflow types to choose from, and quotes can be viewed before running inference.
The author shared a 360-degree orbit MiniMax workflow in ComfyUI, calling it the definitive method they found for multi-angle face consistency. The approach is to first prepare a very solid frontal face image, then have the AI do a 360-degree orbit and capture frames at each important angle; the author claims face consistency reaches 100%, outperforming the search for real photos of real people at different angles.
A user compared six models—Flux Klein 9B Distilled, Flux 3 Klein Distilled, Qwen 2.1 Turbo, Krea 3, Nano Banana 2.1 and Ideogram 4.5—using the same two-image editing prompt.
The author released SyncTile v5.4, a ComfyUI workflow based on MiniMax H3, on Reddit. It splits large-size or long videos into overlapping spatial tiles and temporal windows, and each model call uniformly predicts and blends into the same canvas, thereby avoiding seams and clip stitching, with VRAM usage limited to one region.
The 11 AI video ad generators worth watching in 2026 include Higgsfield, Creatify, Topview, Arcads, HeyGen, Runway, Synthesia, Predis.ai, AdCreative.ai, Pika and OpusClip, covering four task categories: product link to video, AI actor UGC ads, reference ad replication, and batch version testing.
AusBoss released the ComfyUI custom node project ausboss-nodes, adding Image Crop + Rotate + Pad, Load Image + Pad and Load Video nodes, which can load files generated by ComfyUI just like the core Load Image node.
AusBoss released the ComfyUI creative project ausboss-nodes, which supports dragging images from the Assets panel onto node images to load them directly in the classic view.