infinite-creation v1.0.3 released, fixing some bugs and optimizing the categorization of the configuration page. It now supports generating multiple candidate videos from a single storyboard, allowing users to choose the best one for export (equivalent to a gacha/draw feature).
Google Flow can use Veo 3.1 and Gemini Omni to generate videos, but you need to assemble the short film yourself as a sequence of 8-10 second shots. The free version gives 50 credits per day, French subscription starts at 4.99 euros; European users cannot edit imported videos, and Veo 3.1 Qualité does not support reference images. The author calculated the per-minute generation cost for each tier based on official documentation, and noted that Flow is suitable for beginners and trailers, while long-form projects require other tools.
Why it matters: Based on official documentation, this calculates the per-minute generation cost for each Google Flow tier and notes that European users cannot edit imported videos, providing a reference for short film creators evaluating the tool.
Dialects are becoming a differentiated creative lever for AI comic dramas. Four of the top five on Hongguo's trending overall chart are AI comic dramas. 'Our Sword Sect's Beloved Little Junior Sister' used a combination of Henan dialect and adorable-kid cultivation to achieve serialization, leading by a wide margin with five seasons, 1,809 episodes, 23.7 million favorites, and 9.29 million follows. The Northeastern-dialect 'Northeast Stepmother, Straightening Out the Whole Family' ranks seventh on the trending chart. The article argues that the endgame for dialects is not an accent gimmick, but becoming part of the characters' behavioral logic, which is the only way to solidify an IP moat.
Short Drama Study Room believes that the first development cycle of AI short dramas is about to end, and the industry needs to find a new growth curve. AI short drama costs have dropped from 3,000 yuan/minute at the beginning of the year to 800-2,000 yuan/minute. In September, Douyin's native AI short dramas averaged 1,010 new releases per day. From November 5 to 9, 2026, the AI Entertainment (Short Drama) Industry Ecosystem Conference will be held to discuss 2027 industry trends and global market opportunities.
Why it matters: AI short drama costs have dropped to 800-2,000 yuan/minute, overseas traffic acquisition accounts for over 80%, and the industry is entering a cycle transition point.
Dialects are becoming a differentiated creative lever for AI comic dramas. The Henan-dialect cute baby cultivation drama "Our Sword Sect's Beloved Little Junior Sister" leads by a wide margin with 9.29 million viewers, accumulating 23.7 million favorites across five seasons; the Northeastern-dialect "Northeastern Stepmother, Disciplining the Whole Family" ranks seventh on the Hongguo trending overall list. Dialects can help establish character impressions, and AI production reduces trial-and-error costs, but in the long run, they need to upgrade from accent gimmicks to character behavioral logic to become an IP.
AI short dramas are going viral through a proliferating meme-making approach in which 'one meme begets all things.' Memes such as 'not heartburning' and 'sauce-preserved duck' have fissioned into numerous versions on platforms like Hongguo, but formulaic plots are accelerating viewers' indifference to similar content. The article analyzes how the 'not heartburning' meme spread from Hongguo to the entire internet, with related AI dramas reaching about 40 million in popularity on Hongguo and the 'sauce-preserved duck' meme reaching 5 billion views. Themes such as doomsday shelters and 1980s street stalls have also repeatedly proliferated with similar underlying cores. The author believes that AI dramas lacking special innovation will find it difficult to achieve phenomenal traffic again, and core creative memes will become more valuable.
Luma's AI tools for filmmakers (Ray 3.2, Camera Angles, Light Transfer, Face Swap) have matured into production-grade tools that can reduce pre-production time by up to 80%.
Luma Ray 3.2, as a professional alternative to Reface and Remaker, offers native 16-bit HDR export, up to 16 keyframe ordering, and Draft Mode (preview generation 20x faster). Reface restricts commercial use to personal non-commercial, Remaker's commercial terms are unclear, and neither integrates with post-production tools like Premiere or Resolve. Luma's same interface also offers Veo 3.1, Kling 3.0, Seedance, and other models for comparison.
Luma AI leads in a review of 12 AI video face swap tools, with its face swap feature combining the world's first reasoning video model Ray 3.2 with precise editing tools. Face swap tools using temporal consistency perform best in continuous video replacement; DeepSwap supports swapping up to 6 faces simultaneously with 4K HD processing. Magic Hour combines face swapping with lip-syncing and talking photos, with API access for developer integration.
Replicate releases a usage guide for Kling 4.0. The new version supports up to 30 seconds of generation, 720p/1080p output, four aspect ratios (including the new 21:9), and up to 10 keyframes. The keyframe feature can specify intermediate frames, allowing the video to pass through them sequentially in a single continuous shot. The prompt can be up to 26,000 characters.
Why it matters: Kling 4.0 adds continuous shot control with up to 10 keyframes, allowing creators to plan complex actions and transition rhythms accordingly.
On October 24, ByteDance's Jianying system completed business integration, with AI short-drama creation platform Xiaoyunque merging into Jimeng, retaining an independent product entry, with short-drama capabilities integrated into Jimeng and operations handed over to the Jimeng team. This is about seven months after Xiaoyunque launched its short-drama Agent. Meanwhile, Jimeng reduced investment in its app, concentrating resources on the web and overseas Dreamina. The article argues that competition in AI video applications has shifted from feature comparison to operational efficiency testing, and platforms lacking unique workflows will face elimination pressure.
Why it matters: ByteDance's merger of Xiaoyunque into Jimeng shows that competition in AI video applications has shifted from feature comparison to operational efficiency testing, and standalone apps need to prove their own economic value.
The author set up a MiniMAX H3 i2v workflow via ComfyUI on a used RTX 2080 Ti (11GB VRAM). Due to a missing audio_vae resource, errors occurred repeatedly; after filling it in, video generation succeeded.
Kling officially reposted Runware's message, stating that Kling 4.0 is now live on Runware, supporting videos up to 30 seconds at 1080p with native audio. This version supports multiple keyframes for the first frame, intermediate frames, and last frame, with up to 15 inputs per request, covering keyframes, reference images, reference videos, and elements, and can edit existing clips based on prompts. Pricing starts at $0.133 per second of output, with a 15% discount off the regular price for the next 7 days; the model ID is klingai:[email protected].
Why it matters: Kling 4.0 launches via the Runware API, offering parameters such as 30-second 1080p with native audio and multi-keyframe input, which can be compared against existing video generation pipelines.
The core assets of overseas short dramas are shifting from live-action production crews to generative tools, app store rankings and revenue-sharing rules. The article notes that generative video has brought the cost of synthesized short dramas down to the level of tens of thousands of RMB, US live-action output has fallen sharply, and Chinese companies have compressed local production teams into a small supervisory layer; ByteDance's free short drama product PineDrama has accumulated large-scale installs and millions of daily active users within a year by relying on episode unlocks, light gamified retention and heavy advertising, but paid-unlock apps still account for the bulk of in-app revenue.
The article compares five mainstream video generation AIs—Kling, Veo, Runway, Seedance, and Hailuo—pointing out that each service excels in different areas, and that the most popular one is not necessarily suitable for everyone. The author suggests first clarifying what kind of content you want to make, then choosing based on needs such as visual quality, free availability, character motion, and product introduction videos, rather than blindly chasing rankings.
In a music video project, the author tested using a 2x2 storyboard grid generated with GPT image 2 as the reference image for MiniMax reference-to-video. The four panels, in reading order, correspond to the starting frame, two transition frames, and the ending frame. For complex narrative scenes, the adherence effect was better than using only prompts and first/last frames.
The key to turning a text script into an AI video is the workflow, not a single model: split the script by scene, generate a clip of about 4 to 15 seconds for each scene, then assemble them in order and add narration and music. VIBE, an iOS and Android app, integrates models such as Kling, Sora, and Veo, allowing users to generate scene by scene and export an edit on a phone.
Kling officially released Kling 4.0 on October 15, 2026, supporting single generations of 3–30 seconds, more stable large-scale motion and complex camera movements, and up to 10 keyframe inputs. Omni Reference can combine up to 15 reference assets in a single generation, including 10 images, 5 videos totaling no more than 30 seconds, and 7 elements, and prompts can be used to make targeted modifications to existing assets' subject, background, expression, camera position, and style.
Why it matters: The official post reviews Kling 4.0's reference asset, keyframe, and targeted editing capabilities, and offers workflow suggestions from storyboard to delivery.
The official Kling blog compared six CapCut alternatives—Kling AI, DaVinci Resolve, Clipchamp, Filmora, VEED, and VN Video Editor—across platforms, costs, and workflows. Kling 4.0 supports Omni Reference combining up to 15 reference assets, 3–30 second video generation, and up to 10 reference images as keyframes; Kling IMAGE 3.0 Omni can generate images up to 4K.
Kling 4.0 can use Omni Reference to combine up to 15 reference assets, using clear portraits, product images, and the like to guide the reconstruction of faces and details in blurry videos while preserving the original footage's motion, composition, and camera movement. In video editing scenarios, it supports up to 5 input clips (including 1 main video), with a combined duration cap of 30 seconds. The official guidance recommends first determining the cause of blur (motion blur, out of focus, low resolution, compression, low light), then deciding between sharpening, AI super-resolution, or reference-image reconstruction.
The University of Chinese Academy of Sciences, Cardiff University (UK) and the Kling AI team jointly propose JointLight, a controllable relighting method for human videos; the related work was published at SIGGRAPH Asia 2026.
In response to heated discussion on social platforms about "AI short dramas making people dumber," a doctor responded that scrolling short dramas will not cause a medically defined decline in intelligence or organic brain damage; the so-called "getting dumber" is more a temporary reaction of the brain being overloaded after sustained dopamine stimulation, and can recover with sufficient rest. Data shows that in the third quarter of this year, the average daily usage time per user per month on a certain short drama platform was about two and a half hours, and from July to August the average monthly total usage time of free short dramas exceeded 830 billion minutes. The article points out that what truly warrants vigilance is unrestrained addiction, and mentions that the "Measures for the Administration of the Development of Micro Short Dramas" require platforms not to use algorithm models that induce users into addiction.
In response to the claim that "55% of China's token consumption is used for short dramas," Doubao PR head Liu Xing said that only about 10% of ByteDance's total compute is used for video generation model training and inference. That figure comes from an investor relations record disclosed by Chinese Online in June; the statistical scope and methodology were not made public. The so-called US figure of 58% comes from an OpenRouter and a16z report and covers only large language model interactions with OpenAI models on that platform. The two data sets use incomparable statistical scopes, and there is currently no token consumption statistic comparing US and Chinese AI application scenarios under a unified standard.
The Microsoft team proposes Code Video Model, in which a Coding Agent converts a prompt into a Three.js program, unifying geometric layout, object motion, camera trajectories, and event timing into an executable scene representation; the preview video rendered by the program provides spatiotemporal references for the video model.
Why it matters: The Microsoft team uses executable code instead of prompts to control video generation; training-free staged condition scheduling allows shot design to be inspected and modified before generation.
Using the MCP feature of his self-developed Video Builder, the author gave Claude Sonnet 5.5 just one prompt, and it completed the entire rock music video on its own. Sonnet called Krea2 to generate band and scene reference images, wrote lyrics, used YUE2 to compose the music, used MiniMax H3 to generate video, and checked frames scene by scene, revising the prompt and rerunning whenever it found a problem, up to 3 times, all without ever opening ComfyUI.
MOSS-Image in ComfyUI produces a 6-panel storyboard with Japanese descriptions and a table-based 絵コンテ in a single generation. All 18 Japanese descriptions across 3 イメージボード works rendered correctly, while 4 typos appeared across 2 pages of 絵コンテ. Single image generation takes approximately 18–40 seconds on an RTX 3060. The author provides region-separated Prompt writing techniques and a ComfyUI-MOSS-Storyboard custom node workflow.
Why it matters: Local MOSS-Image outputs storyboards with Japanese descriptions and table-based 絵コンテ in a single generation. The region-separated Prompt writing technique and ComfyUI workflow are directly reusable.
A user is asking about the best models, quantization, and workflow solutions for local Text-to-Video and Image-to-Video generation with ComfyUI on an RTX 5070 Ti (16GB GDDR7 VRAM) + Ryzen 5 7600X + 32GB DDR5 setup.
The author has open-sourced Kinoloom, a local video generation tool that runs on ComfyUI. It supports video models such as LTX 2.5, MiniMax H3, and Wan 2.2, as well as image models such as Qwen image 2.1, and can generate videos on consumer-grade GPUs with 8GB of VRAM.
Why it matters: Kinoloom stitches multiple video segments together in latent space and decodes them in a single pass, enabling consumer-grade hardware with 8GB of VRAM to output a seamless 16-second long take with no dropped frames.
The author tested five local MiniMax H3 recipes using the same 15-second three-person alien village scene, with generation times ranging from 95 seconds to 843 seconds.
OmniShowcase proposes a method for generating omnidirectional subject-consistent videos from multi-view images, achieving subject consistency under 360-degree surrounding viewpoints through automatic data synthesis and geometric conditioning. Experiments show it achieves state-of-the-art in subject identity and 3D consistency while maintaining competitive video quality and motion. The research also constructs the Showcase-4K dataset (12 camera motions, 88 trajectory configurations) and the ShowcaseBench evaluation set.
The author ran MiniMax H3 Singularity V2 Pruned INT8 ConvRot R2 with ComfyUI on an RTX 3060 12GB, generating an approximately 10-second video at 1536×864 resolution, 243 frames, and 4-step sampling, taking 608 seconds.
Musk's Grok Imagine has released the first episode of its "Odyssey" series on X, running 7 minutes, with Wonder Studios and director Dev Walker involved in its creation. In July of this year, Musk said Grok Imagine would produce a full-length feature within the year that was faithful to Homer's epic and historically accurate; director Scott Derrickson publicly criticized the project at the time. This release came with no accompanying explanation, and Musk had previously criticized the casting of Nolan's film version.
On the evening of October 25, the author used Meta's AI agent Muse to generate a 30-second vertical video for the self-made "Tianxia Buxian" series. The first three requests produced one 10-second trial and two 30-second versions; the last one was checked and found to be 720×1280, 24fps, 30.000 seconds, with an audio track.
A user is trying to generate loop videos in Minimax using the same first-and-last-frame method, but the motion always "starts" at the 1st second and decelerates in the last 1 second, making it impossible to achieve a seamless constant-speed transition. Taking a 15-second hammering loop as an example, the actual result is 1 second of startup, 13 seconds of steady motion, and 1 second of wind-down before jumping back to the beginning. The user has tried prompts such as "video is a clip of the action segment" with no success, and is now seeking a working workflow, LoRA, or prompt solution.
"Sanxingdui: Future Past" has reached a box office of 1.162 million yuan. The film was released on October 23 and is China's first AI full-process theatrical feature film to obtain a public screening permit from the China Film Administration. It runs 100 minutes, all characters are original digital figures, it contains no digital replication of real actors, and the entire film was completed relying on Bona's self-developed "Boka Film Cloud Studio."
Why it matters: China's first AI full-process theatrical feature film to obtain a public screening permit has passed one million yuan at the box office; its two-year production process and division of labor across 17 job types can be used to assess the degree of industrialization of AI feature films.
A father plans to use ComfyUI on Runpod to make a short film in which his daughter, who loves Game of Thrones, marries her partner in a medieval castle surrounded by family. He says he is new to image generation and workflows, is learning by trial and error, and expects to need DaVinci for editing later, and he is publicly asking for advice.
AIwood Lab found during H3 VBVR testing that loading VBVR did not meet expectations and was even worse; after switching to the older ComfyUI 0.38 environment and comparing with 0.42, noticeable differences appeared in scenes such as the chef.