The AI manga "The Pasta Connecting Tamaki and Hiyori" has updated with Chapter 7; the theme of this chapter is clam white wine pasta (ボンゴレビアンコ). The author mentions that an American cream-flavored version will appear in later content.
The author has released PixText, a model with fewer than 1B parameters (about 0.95B) for fixing garbled text and Chinese characters in generated images. The model is trained directly in pixel space, requires no text encoder or VAE, and uses MeanFlow to complete text repair with only 4 sampling steps. In ComfyUI, users can select a region and enter the correct text, or first automatically detect text and then adjust the suggested boxes; a comparison node shows the before-and-after effect for each region.
Why it matters: To address the problem of small text in generated images being blurred by VAE compression, the author provides a pixel-space repair model and a ComfyUI box-selection text-editing workflow.
ByteDance's Volcano Engine has released Seedance 2.5 4K, calling it the strongest model in the Seedance 2.5 generation, with specialized training to learn character expressions, subtle movements and image details. The model supports a Draft sample mode, allowing users to first verify storyboards and polish shots with 480P samples before generating the 4K final cut, and it features 10-bit color depth; it is currently available only for trial by some enterprise customers, and some film and television production and post-production companies at home and abroad have already tried it.
Why it matters: Seedance 2.5 4K splits sample verification and the 4K final cut into two stages, allowing film and television teams to judge how to trade off image quality and cost accordingly.
A ComfyUI user is looking for a workflow that can generate several-minute-long NSFW videos; most existing workflows max out at about 10 seconds, or require a prompt every 5 seconds to extend. The user also encounters object permanence issues: after a character in the frame is briefly occluded, details such as clothing are incorrectly restored. The user uses Vast.ai hardware, so compute power is not a constraint.
A creator uses a workflow connecting Claude Code to ComfyUI MCP, mainly relying on Qwen Image 2.1 for drafts, design, atmosphere and story, then moving to H3, and plans to try Seedance 2.5 in projects. He has never used paid models and is unsure about workflows, so he asks the community: what tools and models should be chosen when the budget is sufficient, which is the best image model, and what everyone's workflows are like.
At a short drama overseas expansion industry event, DataEye disclosed the January-September 2026 overseas micro short drama app rankings: Kunlun Wanwei's FreeReels topped the list with over 230 million downloads, Maple Leaf Interactive's ReelShort ranked second with over 190 million, and Maiya's NetShort ranked third with over 180 million.
Before procuring an AI renderer, buyers should check whether the vendor publicly documents input formats, host versions, quota consumption, updates and failure recovery, rather than only looking at rendering results. The article suggests testing item by item with six questions: what can be input, what will change, what is output, how billing works, how data is stored, and what updates break; and having a colleague not involved in selection run a complete workflow using only the public documentation. Products that rely on sales talk, old videos or private support cannot become a reusable studio standard.
ArchiGen points out that "free" in free AI rendering solutions only indicates price and does not mean the account type, project use and commercial output are licensed. Buyers need to check eligibility terms, commercial-use terms, output rights, quota reset and export deletion rules, and keep dated copies of the terms. Under free quotas, previews, failed tasks, variants and upscaling may share the same metering unit, and the resolution, watermarks and metadata of exported files also need to be confirmed before delivery.
A user asks for help in the ComfyUI community: how to put different people into the same image without losing their respective facial identities, and have them pose and do other actions together. The user has tried krea 2 and flux Klein 9 distilled models with a basic image editing workflow, with hardware of an AMD 6570xt, 12GB VRAM and 32GB system memory, and hopes someone can recommend a feasible workflow.
pyvideotrans v4.17 released, adding the Audio.cpp speech recognition channel, and adding noise reduction and vocal separation models based on audio.cpp. It also fixes the nemotronasr repeated download issue, adds the Opper translation channel, and adds a 'further simplify' option to subtitle translation.
Why it matters: The new version adds the Audio.cpp speech recognition channel and noise reduction models, and optimizes the subtitle translation simplification option, suitable for evaluating upgrades to the video translation and dubbing workflow.
Promptsref has announced this week's three hottest Midjourney SREF styles: retro Japanese animation illustration, pink-gold dreamy surreal fashion, and classical engraving-style hand-colored prints. The three correspond respectively to late-night OVA nostalgia, commercial fashion editorial, and museum print texture, and can all be quickly tested via --sref and reused for series creation.
The Liaoning Provincial Radio and Television Bureau held a discussion with Liaoning Advertising Vocational College on October 22, reaching consensus on advancing the high-quality development of Liaoning's micro short drama industry and establishing a government-industry-enterprise-college-research collaborative mechanism. The provincial bureau proposed implementing a "four ones reserve project" for micro short dramas to shore up shortcomings in the industry chain, while the college proposed cooperation and co-construction items based on its fully closed-loop industry-education integration system of "training—practical training—creation—broadcast—marketization." Both sides will deepen the competition brand and build a Liaoning-exclusive, nationally renowned cultural audiovisual IP.
The author used MOSS-Image to test 20 anime and manga styles, covering 1990s cel animation, Shonen Jump covers, Ken Sugimori watercolor, sakuga and kanji typography, and also included a ComfyUI workflow.
Kling 4.0 is now available on Pollo AI, supporting keyframe-controlled sequences, multi-reference material-guided creation, and video editing. Pollo AI's Ultra users can get 10 days of unlimited use of Kling 4.0.
A user is asking for help with a ComfyUI workflow, hoping to use one large prompt to automatically split 45 shots (two people talking in the same scene, multiple camera angles, each clip with different audio) and generate them segment by segment, then import them into DaVinci Resolve for editing. Currently he can only manually prompt and render segment by segment at 3-5 seconds each.
Novae has completed her first full vertical short drama and has published it on platforms including IG, FB, and YouTube, while also using Astra to generate a cinematic book video in about 15 minutes. Cassie is producing a visual novel and has released a new film, and her Kickstarter is progressing smoothly. The two are also trying a Blender-to-Minimax H3 video pipeline to reduce costs, and will share it in a livestream at 10 a.m. Pacific Time on Friday.
"Sanxingdui: Future Past," released on October 23, is the first AI hyper-realistic film in China to obtain a theatrical public screening license. Before release, fewer than 6,000 people marked "want to see" on Maoyan, and Beacon Professional Edition predicted a total box office of 2.169 million yuan, while the production cost was in the tens of millions of yuan.
AnyGameRenderer is a video-to-video generative renderer that can turn placeholder visuals such as gray-box and colored geometric shapes produced by AI agents into high-quality footage with characters, motion details, and effects, while keeping the layout, camera, and the position and timing of each entity unchanged.
Kijai's MOSS-Image Edit LoRA support has been merged into the latest ComfyUI Portable, giving MOSS-Image Edit model capabilities and ControlNet-like basic functions. Supporting releases also include the CharRef v2 character reference node, a 9.4Gb Dirty MOSS realistic fine-tune, an animation style LoRA, and a library of 500 usable prompts with example images.
"Sanxingdui: Future Past" opens in theaters nationwide on October 23. As the first AI hyper-realistic film in China to obtain a public screening license, before release it had fewer than 6,000 "want to see" marks on Maoyan and about 17,600 on Taopiaopiao; total presales across the two platforms were about 144,000 yuan, and Beacon predicted a first-day box office of 311,000 yuan and a total box office of 2.169 million yuan, while the film's cost was in the tens of millions of yuan.
Why it matters: Using the cold presales of the first AI theatrical film as an entry point, it compares IP revenue sharing and satellite TV paths, and sorts out the differences among model vendors entering the film production process.
Kling officially released a behind-the-scenes breakdown of KING CARD, introducing that the film received 5 million views on social platforms and explaining how Kling 4.0 was used to complete shots such as flying cards, a newspaper train passing through the city, and environmental transformation.
Why it matters: Kling officially breaks down the production process of KING CARD, showing how dynamic motion, multiple keyframes, and Omni Reference work together in complex shots.
After comparing about 15 image generation AIs, VN.ai found that Midjourney had the highest proportion of generated images at risk of infringement, about 37%, and that it had improved little about a year and a half after being sued by two film companies. NTV HR Research Institute used generative AI to produce 3 recruitment commercials in different styles for the Imizu City fire brigade in Toyama Prefecture, including a 32-second horizontal live-action style, a 37-second vertical animation, and a 35-second horizontal live-action.
Kling 4.0 is now available on idea.ai. According to idea.ai, this version can generate a 30-second scene from a single prompt, maintain character consistency across shots, and support switching between live-action, anime, and 3D styles.
Why it matters: Kling 4.0 launches on idea.ai, supporting 30-second scene generation from a single prompt and cross-shot character consistency, and can be evaluated against existing output pipelines.
To address the problem that ComfyUI users cannot remember the effects of lighting prompts, the author created a custom node where clicking lets you select lighting expression terms such as backlight, spotlight, God Rays, particles, low-key, and specular reflection, and hovering the mouse also displays text explanations. In the example, after overlaying these terms on an armored character, the image changes from flat to a heavy, solemn atmosphere. The author also notes that newer-generation models such as Krea2 already have fairly high inherent quality, but this node can be used to precisely hit the desired lighting effect.
Verus-Insight breaks down artifact detection for AI-generated video into three stages: inspection planning, evidence verification, and evidence aggregation, and uses human-anchored full-pipeline supervision and progressive capability training to improve fine-grained artifact recognition and video-level detection. The team also released the Verus-Bench benchmark with fine-grained human artifact annotations. Experiments show that the model has a clear improvement in recovering human-perceivable artifacts without increasing false artifact assertions.
A ComfyUI beginner is seeking workflow advice, hoping to use Trellis to generate a 3D model from a 2D photo of a person to make a statue of a friend, with results benchmarked against Bambu Lab's "Make My Statue" studio on MakerWorld. His current results with Trellis are far below Bambu's, but he does not want to pay for every attempt and hopes to do it locally.
A developer is building LoRA Studio, a local browser interface for dataset preparation, caption editing and managing LoRA training via AI Toolkit, currently focused on training Krea 2 on 12 GB VRAM GPUs, with 8 GB as a later goal.
A beginner asks whether their configuration of an i5-13600K, AMD 6700 XT with 12 GB VRAM and 32 GB DDR5 memory can be used for video generation practice. They also want to know whether they should use video editing software to splice short clips into a long video, and whether they should upgrade directly to an AMD 7900 XTX with 24 GB VRAM, or whether that card would still be insufficient with 32 GB of memory to make video generation faster.
VIBE is an AI video generation app for iOS and Android that turns real estate photos into 5-to-10-second clips for each room via image-to-video, then splices them into a property tour. The app supports models including Kling 3, Veo 3.1, Seedance 1.5 Pro and Luma Ray 3.2, outputs 16:9 and 9:16, and Seedance Pro Fast can be tried for free.
Hongguo Short Drama announced a phased shutdown of its coin feature, no longer offering it to new users from November 1 and fully discontinuing it by December 31; existing coin and cash balance benefits are unaffected. WeChat's Green Bubble Short Drama, meanwhile, is competing for quality new dramas in the fourth quarter with comprehensive revenue incentives of up to 100%.
Haishen Hyacinth shared a video production workflow combining Claude Opus 5.5 and Seedance 2.5: first use Opus 5.5 to create dynamic storyboards, then produce final-effect keyframes, then use the keyframes as references and hand them to Seedance 2.5 to generate video.
Multimodal models such as MiniMax H3 and Seedance are being integrated into real content workflows in gaming, e-commerce and other fields. A large game company used them to shorten asset production from several days to several hours, and monthly costs fell from several million yuan to several hundred thousand yuan. H3 outputs video in 768p and 2K modes; 2K API calls cost about 0.8 yuan per second, and local deployment costs less. MINIMAX-W revenue in the first half of the year was $117 million, with open platform and enterprise service revenue up 703.1% year over year.
The open-source creation tool released v1.6.9.2; after closing short drama production, the project entry in the sidebar and command palette is now hidden accordingly. This version fixes the proxy address pointer parameter in the channel migration test, restores compilation of the Go test package, and has no database schema changes.
This tutorial compiles a short demo of the Singularity-V2 bundle from Liu Yue's tech blog, explaining that the bundle defaults to 4-step iteration. The short video actually demonstrates three input methods: text-to-video, first/last frames, and all-purpose reference. The author states the demo device is an RTX 4060 laptop, with a single generation taking approximately 100 seconds; this duration corresponds only to the conditions of this demo. To view node connections, you can access the ComfyUI workflow via custom port 8187; the port should be confirmed from the local startup log.
A team from CUHK MMLab, ByteDance and HKUST proposed DuoMatching, which lets a video generator accept both joint-distribution supervision and frame-level marginal-distribution supervision, introducing image priors during training without adding inference computation. Under the per-frame 2-step setting, the VBench total score rose from 80.67 to 83.53, the overall preference rate against all four baselines exceeded 80%, and the dynamic degree changed slightly from 94.44 to 93.61. The paper is public on arXiv, and the code and model weights are released together.
The author released a free, open-source ComfyUI node that improves diversity within the same batch by coupling initial noise across multiple images, without extra sampling steps or models—only the noise node needs to be replaced. The node supports SD1.5, SDXL and FLUX.1-schnell, comes with an example workflow, and the code is in the coupled initialization project on GitHub. The author says this is their first ComfyUI release and hopes for feedback.
A 102-page survey proposes world modeling as a unified perspective, organizing research on video foundation models and video world models into four operations: read, write, share and interact. Read infers world properties from observed video; write synthesizes video under specified or learned world conditions; share integrates understanding and generation with common modeling mechanisms; interact predicts and updates world states under actions and evolving observations.