TVB starts shooting short dramas
TVB starts shooting short dramas.
TVB starts shooting short dramas.
CMG National Medium and Short Drama Conference releases more than 30 medium and short dramas.
As of October 19, 2026, CapCut has no official MCP server or public editing API. Community servers (VectCutAPI and its wrappers, and independent projects) can only generate CapCut desktop drafts by writing draft JSON files, and export still requires manual action in CapCut. These servers all run locally and require CapCut and the AI agent to be installed on the same machine; hosted MCP servers complete editing and rendering in the cloud and return video files.
A user on the Reddit comfyui subreddit is asking for help reaching 2MP with Krea2 turbo via an upscaling workflow, but multiple versions tried all produced random issues that degraded image quality. They have tried samplers such as deis, bongtangent, euler, and beta, and ask whether only Krea2 Raw can achieve this, and whether LoRAs should be used only in the first pass, only in the second pass, or in both.
A user is training a LoRA for Moss-Image base with AI-Toolkit on an RTX 4070 with 16GB of RAM, and even after updating to the latest version, all configurations run out of memory. The user mentions an AcademiaSD post claiming training is possible with 6GB of VRAM, but is unfamiliar with that tool and asks whether AI-Toolkit can achieve the same.
A user has been using Runpod with Minimax H3 for video generation in ComfyUI for about a month, and due to a recent price increase is looking for cheaper alternatives that can also serve as backups. A common recommendation, vast.ai, is slightly cheaper but has poor availability of 5090-class GPUs; smaller platforms like glows.ai quote lower prices but lack review information. The user wants alternatives that allow NSFW and ideally offer a direct ComfyUI deployment template.
A ComfyUI beginner asks on Reddit for a complete character replacement workflow: input their own AI influencer reference image and a reference video, and generate a video that keeps the original video's motion and expressions but swaps in their own character, while maintaining character consistency. They also ask about the actual cost of generating 20-30 second long clips, whether free models are available, and whether ComfyUI is the cheapest option or whether alternatives exist.
A user wants to use an existing RTX 4080 and a newly acquired RTX 3090 simultaneously for ComfyUI, but neither the motherboard nor the case can fit two cards, so they are considering converting the 3090 into an external GPU via Thunderbolt. They want to know whether an external setup can reach about 70% of the performance of a PCIe x8 slot.
A ComfyUI beginner posted on Reddit asking how to achieve the realistic AI character replacement effect posted by a certain account; despite having an AMD Radeon RX 9070 XT and 32GB VRAM, they still couldn't get the reference video workflow to run. They said they couldn't find relevant tutorials on YouTube and hoped the community could recommend a usable workflow.
A user asked whether a configuration of i5 14600KF + 64GB DDR4 + 5060 Ti 16GB can run Wan2 14B, Flux 2 dev, or heavier models on ComfyUI, and said that long runtimes are not a problem. The user is also considering using the dual GPU slots on the MSI Z790P motherboard (PCIe 5.0 x16 + PCIe 4.0 x4) to add another 5060 Ti 16GB and split heavy models across the two GPUs.
A user asked for help in the comfyui subreddit on Reddit: they run local ComfyUI in Docker with an RTX 2070 SUPER 8GB GPU, and can already run an audio deep synthesis workflow based on a target vocal recording. They want to go further into video AI face swapping (video only, or video plus audio), and guess they need to combine the video with the audio deep synthesis result for lip sync.
A user asked for help in the ComfyUI community, wanting to recreate a type of funny video popular in 2024: people suddenly twisting in weird ways, with objects randomly intruding, and said these videos are performed by real people imitating. They posted several examples and asked which model was used at the time and what prompts are needed to achieve this effect.
How many? 21 million likes? Am I reading that right?
A user on the Reddit StableDiffusion subreddit posted asking for recommendations of open-source image models runnable locally as of October 2026, targeting Hollywood-grade cinematic realistic stills with natural skin texture, realistic lighting, and professional color grading, without the typical AI look. The hardware limit is an RTX 3060 with 12GB VRAM, and the poster also asked for available fine-tuned models and ComfyUI workflows.
The community conducted a hundred-group in-depth review of Moss Image 4, comparing it head-to-head with Qwen Image 2.1 and Flux3 Klein. The review also covers the T8 open-source prompt enhancement node and the community's 8-step LoRA acceleration solution.
Creator KITAcore used ChatGPT and image generation AI to create the otherworld hero "Takeshi"—an ordinary company employee summoned to another world who insists on exchanging business cards—and opened a dedicated website for him, "Otherworld Hero Takeshi | Official War Fundraising Office." The site shows a subjugation progress of 0%, and the first round of merchandise, phone cases, is already on sale. The character first appeared in a September 2025 article about newcomers carrying business cards in cardboard boxes, and has since quietly appeared in multiple notes and comment sections.
DeepSeek-V4.1-Flash-1017 reproduced a 3D voxel scene of the new Tesla Roadster climbing a hill plus a flying stunt for $0.89, while GPT-6 Luna cost $1.07, about 17% lower in cost. The official announcement claims major improvements on agentic tasks such as Terminal-Bench, stronger vision and coding capabilities, and open-source MIT weights that can run locally or via API. The voxel results of the two are close, and DeepSeek offers better value for money.
DeepSeek-V4.1-Flash-1017 has been updated and released, with improved benchmark scores and shorter outputs, at an unchanged price. In testing, it used a single prompt to generate, purely in code, a 2-minute 1920×1080 60fps cinematic video, with the soundtrack and every frame generated by code; the result is close to GPT-6 Astra and weaker than Opus-5.5.
Xiaohongshu is absolute crap, I'm so mad 😤
A user asked Reddit for help generating both video and images on a Lenovo Legion 7 16IRX9 and wanted operating guidance. They had used Automatic1111 before but had forgotten how to configure it, and had heard the tool is outdated, so they asked whether they should switch to Flux.
A Reddit user posted in the comfyui subreddit asking for help learning ComfyUI to make realistic and consistent AI models, hoping for YouTuber or playlist recommendations as beginner resources. They said they are willing to invest a lot of time and want to truly understand the principles rather than copy others' workflows.
This AI illustration morning share takes "starry sky in the dead of night" as its theme, using GPT Image 2.5 to draw 4 works under the same moonlight yet each with its own expression and worldview. The author also recommends an article by Funfun Shokudo and includes a daily healing message.
The Runway plugin is now connected to ChatGPT, letting users generate product videos, stylized photos, or complete ads in natural language without an API key. Google is broadly opening Gemini 4 Argon, with a limited-time API price of $2/$10 per 1M tokens. Qwen-Image-2.1 adds multi-angle 3D camera control, and Qwen3.8-Flash-Next supports one-click installation on Windows / Linux.
ByteDance's overseas short drama app PineDrama surpassed 100 million downloads in 9 months, but it has not replicated Hongguo's domineering-CEO formula. The article analyzes the distance between it and an overseas version of Hongguo and explores the differences in ByteDance's path for short dramas going overseas.
T8 has open-sourced a PE node dedicated to Flux3 Klein character ID consistency, used to solve the model's character identity preservation problem. The original text also lists supporting resources such as the T8 cloud mirror, the ComfyUI integrated package V10, and several trainer mirrors.
A user asked on Reddit whether renting RunPod's GPU/VRAM with the free ComfyUI to run the MiniMax H3 open-source model is the cheapest option for making high-quality AI videos. The user believes that although some people complain ComfyUI is restrictive, it allows fine control over details and is better than Higgsfield, and mentioned that some people connect ComfyUI with Copilot to lower the learning curve.
A user on Reddit asked whether a high-quality pose concept LoRA can be trained with just one reference image, to reproduce in the Anima model poses, body curves, camera angles and compositions that are hard to describe with Danbooru tags. The user said OpenPose cannot accurately extract skeletons and preserve pose details, and that they have never trained a LoRA, and wanted to know whether other alternatives exist.
The real transformation in vertical short dramas happens behind the camera, not in front of it. In July 2026, NBC's "The Tonight Show" released a 10-episode parody short drama, "Who Billionaire? Me Billionaire.," spoofing the "secret billionaire" genre.
A user reported that after a fresh ComfyUI installation, saved images no longer carry metadata, especially prompt information. The user asked whether this is intentional or a missed setting, and noted that it affects subsequent regeneration of images with modifications.
AusBoss releases ComfyUI creation project ausboss-nodes, with the changelog kept on GitHub. The project offers a smaller download size.
An Indian AI alliance claims its solution can cut film and television production costs by 75%. This content comes from a paid-subscription report by C21Media; the main text does not disclose further technical details, participating companies or implementation timelines.
A user on Reddit asked whether there is a Video2Video model similar to DLSS 5 that can enhance CGI videos to be closer to photorealistic, rather than just conventional upscaling, denoising, and frame interpolation. The user has a 7900XTX with 24GB of VRAM and said they are willing to rent compute if a larger model is needed.
A user loaded MalRey's character LoRA in Comfy's default t2v H3 template, but the video output was not affected at all. The user also tried adding the character name to the prompt, which still did not work, so they posted to ask whether this is related to how the prompt is written.
Tongyi MAI has launched the MAI-Gym-9B model. Currently, the model has not been deployed by any Inference Provider, and users can apply to service providers for support. The model's download count last month was not tracked, and the README documentation is empty.
Outrageous!!! This is actually not AI. Can anyone still be inside this car? What does it feel like to flip around inside it?
Holy crap, Tesla's new sports car is insane—it can actually fly briefly and adjust its attitude in the air. In the demo it can jump up, flip, and land normally.
An author who writes harem and light-novel web fiction has been trying for several days on ComfyUI with an RTX 2060 6GB VRAM, 32GB RAM, and Illustrious, wanting to train style and character LoRAs to produce consistent marketing illustrations, but has only generated garbage results. They posted asking for tips, and also need to make NSFW images of characters because of the harem genre.
Flux3 Klein has been open-sourced, and the most comprehensive comparative review with Qwen Image 2.1 has been conducted. Related supporting resources have also been released simultaneously, including the Flux.2 Klein full-feature trainer image, the Qwen Image Ai toolkit full version trainer, as well as the T8 ComfyUI integrated package V10 and the Zhenzhen integrated package V4.6B.
A ComfyUI user, pairing Krea 2 RAW with an identity LoRA, is seeking recommendations for a realism LoRA that can make AI images realistic to the point of near-photos, focusing on realistic skin texture, visible pores, facial details, natural lighting and fabric texture. The user hopes for recommendations of LoRAs that do not make skin look plastic and do not change the model's facial identity, and requests LoRA names, links and recommended weights, while also asking which model architectures they are compatible with.
A creator asked on Reddit about wanting to turn AI-generated anime images into 3–5 second short videos or GIFs, but their local computer cannot run heavy models such as Wan, LTX and Hunyuan, so they hope to find simpler cloud-based alternative tools. Their requirements include support for adult-oriented anime illustrations, reasonable pricing, and no need to spend hours debugging ComfyUI workflows.