The front line of text-to-video and image-to-video: tracking model releases, capability boundaries, price changes, and transformations in film and television creation.
344 picksLast 30 days 240 itemsCollected 1,725 items
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.
ComfyUI launched FLUX 3 Max, a latency-optimized version of FLUX 3 Video that can generate 10-second 720/768p clips with native audio in about 12 seconds. The official release also provides links to two workflows, image-to-video and text-to-video, and lists use cases such as comparing multiple candidates, previewing motion while tuning parameters, batch-producing ad variants, and keyframe-guided storyboards.
Why it matters: ComfyUI integrates FLUX 3 Max, offering a latency-optimized version that produces 10-second 720/768p clips with native audio in about 12 seconds, along with entry points for image-to-video and text-to-video workflows.
Kling's official account teased the next-generation video generation model with #Kling4, without giving specific parameters in the post. The Morphic release information it cited says Kling 4.0 supports native 30-second generation, up to 10 keyframes, 15 references, on-screen text consistency, and multiple languages, and offers 4K and 10-bit HDR, with a 10% discount for generations on Morphic before November 20.
Why it matters: Kling officially teases Kling 4; cited information indicates it supports native 30-second generation and 4K 10-bit HDR, which can be used to gauge its long-shot output capability.
Kling announces that the KLING 4.0 API is officially open, allowing developers to integrate Kling 4.0 into their own products and workflows. The original text does not disclose specific capability parameters, pricing, or integration details.
Kling 4.0 is now live on fal. The official description calls it a new-generation native multimodal video model with improvements in complex motion stability, character and scene reference consistency, and cinematic narrative control. Currently, the model is only available through fal UX and fal Agent, with API access coming soon.
Why it matters: Kling 4.0 launches on fal, with the official announcement citing three capability changes: complex motion stability, character and scene reference consistency, and narrative control.
Kling announces that the KLING 4.0 API is open for applications, allowing developers to integrate Kling 4.0 into their own products and workflows. The original text does not disclose specific interface capabilities, pricing, or opening time.
Tech Planet reports that domestically produced AI Korean, Japanese, and American dramas are going viral on Douyin and Xiaohongshu; among them, the AI Korean drama "Beautiful Malice" surpassed 180 million views after six episodes, gained more than 1.7 million Douyin followers, and monetized millions of yuan. The show's per-episode production cost is about 15,000 to 20,000 yuan, and it mainly monetizes through incubating story accounts for advertising rather than platform revenue sharing. In the first half of 2026, about 221,900 AI short dramas launched on Douyin, with about 1,055 surpassing 100 million views, a hit rate of about 0.47%.
Why it matters: Using cases such as "Beautiful Malice," the piece breaks down the account incubation and advertising monetization path of domestic AI Japanese and Korean dramas, and provides industry data on hit rates and falling costs.
Xiaguang Think Tank released a report on Chinese AI short dramas going overseas, noting that AI short dramas account for 95% of production but only 4% of views, and the industry is shifting from competing on volume to competing on quality. The report shows that the production cost per AI short drama has dropped from 500,000–1.5 million yuan to 50,000–240,000 yuan, with efficiency improving 60–100 times, but overseas traffic acquisition costs account for about 50% of total overseas costs, and the customer acquisition cost per user in North America reaches $5.28. The competitive focus is shifting from content production to traffic acquisition, and compliance has become an entry threshold.
Why it matters: AI short dramas account for 95% of production but only 4% of views; the report reveals a mismatch between capacity and traffic, providing a reference for overseas players to shift from competing on volume to competing on quality.
The author compared the rendering results of Wan 2.2 under two tools, Loomance and WanGP, and found that the 4-step lightning mode is about 10 times faster but has more errors, while the 30-step full sampling is cleaner. In the six-dog scene, WanGP had 3/5 instances of character melting or merging, while Loomance had 0/5; in high-resolution single-character scenes, the two are close. Choose as needed: use WanGP for rapid changes or low VRAM, use full sampling to maintain character consistency.
Why it matters: A practical comparison of the same Wan 2.2 model under two local tools offers a reference for creators to choose between speed and fidelity.
Tencent Hunyuan open-sourced the interactive world model Hy-Nomas on October 21, 2026, releasing weights, inference code, and training and data annotation code, along with an online demo and the Hy-Nomas-Bench benchmark.
Why it matters: Tencent Hunyuan open-sources an interactive world model in which text and camera trajectories can drive long video generation segment by segment, with a benchmark and single-GPU configuration.
FastVideo released FastH3 Omni, a reference-conditioned variant of MiniMax H3; a prompt plus ordered image, video and audio references can generate clips with synchronized sound, requiring only 8 transformer evaluations instead of 50 for base H3.
Why it matters: FastH3 Omni unifies reference images, video and audio into a single generation pipeline, and provides an item-by-item time comparison with base H3 to help assess local deployment costs.
ComfyUI announced support for Hao Ai lab's FastH3 OmniRef, a model aimed at fast, high-quality video and audio generation. OmniRef enables support for multimodal reference materials, so creators and teams no longer rely solely on text prompts to control generated content; the model can run locally, allowing teams to generate and retain materials on their own hardware without relying on hosted APIs. The official announcement also provides a cloud entry point and a Hugging Face open-weights link.
Why it matters: ComfyUI integrates FastH3 OmniRef, supporting multimodal reference materials and local operation, so creators can assess building their own video production pipeline accordingly.
The FastVideo team at Hao AI Lab has open-sourced FastH3 OmniRef, which generates video with synchronized audio from a prompt plus reference images, short video clips, or sound. The model needs only 8 steps instead of 50, is up to 11.9 times faster than the base MiniMax-H3, has entered the ComfyUI 0.42.2 workflow templates, and has open weights and training code.
Why it matters: FastH3 OmniRef is released as open weights, produces video in 8 steps, and supports day-one ComfyUI integration, so creators can use it to evaluate local reference-to-video pipelines.
FastVideo released ComfyUI-usable weights for the final version of FastH3 Omni, providing two diffusion models, BF16 and INT8 ConvRot, which can generate video and synchronized audio from prompts and ordered image, video, and audio references.
Why it matters: FastH3 Omni releases ComfyUI-usable BF16 and INT8 weights and provides the way to connect reference images, video, and audio, making it easier for creators to evaluate local video production pipelines.
The author open-sourced the ComfyUI custom node suite ComfyUI-ZeroCopy-Continuity, which uses ZeroCopyVRAMPin to keep DiT weights resident in VRAM and ZeroCopyLatentHandoff to pass the previous shot's latent directly on cuda:0, eliminating the VAE decode-then-encode round-trip.
Why it matters: The author open-sourced a set of ComfyUI nodes that replace VAE round-trips with VRAM residency and direct latent transfer, reducing the transition overhead in multi-shot rendering from 14.2s to 0.4s.
ComfyUI-RMBG releases v3.3.0, adding a Canvas interactive multi-layer compositing node, temporal video matting based on BRIA AI VRMBG-3.0, FeyNoBg fine matting for hair strands and glass, and YOLO26-seg and YOLO11-seg segmentation support.
Why it matters: This update integrates matting, detection, and compositing into the same set of ComfyUI nodes and adds temporal video matting, allowing evaluation of replacement costs against existing production workflows.
The author breaks down an AI face-swap video workflow from raw footage to finished video, using ten folders numbered 00 to 09 to divide stages such as shot planning, reference images, generation packages, frame alignment, final compositing, and quality control. In the generation stage, the pipeline is run twice on ComfyUI with three NVIDIA DGX Sparks, first a 1280×736 draft and then a 1920×1088 refinement; an 8-second 1080p-level clip takes about 1 hour.
Why it matters: The author reviews the full workflow of a face-swap video with 54 shots and 8163 frames, focusing on easily overlooked engineering steps such as frame alignment and version management.
The author introduces a new combined workflow for MiniMax H3 in ComfyUI, using the lightweight timeline node obvpm-timeline together with RefMod for fixed characters to complete generation, backward extension, forward insertion, and split insertion within the canvas. The article says the timeline connects directly to the latent data of each shot, avoiding the quality loss and audio-video misalignment caused by MP4 concatenation, and includes links to the workflow, Colab launch code, and node repository.
Why it matters: The author tested combining obvpm-timeline with RefMod, demonstrating continuous short drama production with forward/backward extension and split insertion within ComfyUI.
VX Studio founder Shaurya Gupta uses a 90-episode vertical drama with about 1600 generations as an example to explain that AI video engines have no memory, so the same character across shots shows drift in face shape, age, and height, and by the fourth consecutive extension it already looks like a different actor.
Why it matters: The founder of an AI short-drama production platform reviews hands-on experience from a 90-episode vertical drama, listing point by point the failure points of character consistency, shot extension, and wide-shot distortion, along with countermeasures.
Frank Houbre gives a production method for keeping the same character across shots: first make a front-facing main image plus two or three angle references, then use the model's built-in character reference entry point rather than pure text prompts, and finally break down the storyboard by shot risk.
Why it matters: The author horizontally compares the character reference entry points of Kling, Seedance, Flow, and Runway according to official documentation, and provides a reusable storyboard and reference image preparation process.