aivideo.zone AI video news
Model releases, open weights, ComfyUI updates and API changes, newest first. Each one says where it came from.
After Sora: the AI video APIs still taking requests
OpenAI switched off the Sora 2 API on 24 September with no replacement named. Four AI video APIs a Sora integration can move to — xAI, Google, MiniMax and Kling — and what each one asks of you.
One scene, three prompt formats
A screenwriter's guide to the three ways AI video models want a scene written: LTX reads a flowing paragraph, MiniMax H3 a tagged shot list with timed cuts, Kling a timeline pinned to keyframes. The same scene in each.
Editing AI video in ComfyUI
ComfyUI is no longer only where AI video gets generated. September's releases added nodes to join clips, loop over them, erase objects, edit footage with a prompt and grade in film colour spaces. What each one does.
Kling 4.0 vs Kling 3.0: what actually changes
Kling 4.0 doubles the clip length to 30 seconds, adds up to ten keyframes, stereo sound and 10-bit HDR, and more than doubles the references one AI video can take. Kling's own comparison, line by line.
LTX-2.5 gets fourteen LoRAs and HDR
Lightricks spent September turning LTX-2.5 into a post-production kit: fourteen LoRAs for jobs like mattes, deblurring and day-for-night, long AI video in overlapping windows, and an SDR-to-HDR pass that writes EXR.
Writing a shot list for Kling 4.0's ten keyframes
Kling 4.0 lets one 30-second AI video pass through up to ten pinned keyframes, with a prompt of up to 8,000 tokens to carry it between them. How Kling's guide says to plan that, written up for screenwriters.
Kling 4.0: 30 seconds, ten keyframes
Kling AI announced Kling 4.0 on 30 September: native AI video of up to 30 seconds, up to ten keyframes, stereo sound and up to fifteen references. A Flash version is in limited early access; the full model comes in October.
ComfyUI 0.38: Seedance drafts, H3 ControlNet
ComfyUI 0.38 brings Fun-ControlNet-Union 2.0 to MiniMax H3, a Draft mode for ByteDance's Seedance 2.5, film colour spaces for grading and a lighter install — and drops its Sora nodes now that OpenAI's AI video API is gone.
The Sora 2 API shuts down
OpenAI removed its Videos API and every sora-2 and sora-2-pro model on 24 September, six months after giving notice, and named no successor. It ends OpenAI's AI video offer for developers; the Sora app closed in April.
ComfyUI 0.36: join, loop, edit
ComfyUI 0.36 adds a core node to join clips and a way to loop part of a graph, plus partner nodes for BFL's Flux Video Edit, Bria's Video Eraser and Pruna's P-Video-2 — AI video editing without leaving the workflow.
A Grok Imagine image model retires
The still that starts an image-to-video AI video clip often comes from an image model. xAI is retiring one of Grok Imagine's on 2 November and moving its requests to Image 2.0, which gained an automatic quality setting last week.
Gemini Omni 1.1 Flash is generally available, now up to 4K
Google made gemini-omni-1.1-flash generally available in the Gemini API on 27 August, adding AI video extension, interpolation between two images and output up to 4K. The preview it replaces shuts down on 30 September.
LTX-2.5 or MiniMax H3?
Two open-weights AI video models that generate sound with the picture came out within a fortnight of each other. LTX-2.5 and MiniMax H3 compared on size, output, references and the licence terms that decide who may run them.
ComfyUI 0.32: LTX-2.5 and Grok Imagine
ComfyUI 0.32 shipped on 11 August with support for LTX-2.5, Lightricks' new open AI video model, partner nodes for Grok Imagine Image 2.0 and Qwen-Image 3.0 Pro, and a higher minimum PyTorch.
LTX-2.5 is out, with open weights
Lightricks released LTX-2.5 on 11 August: a 22-billion-parameter open-weights AI video model that generates sound with the picture, cuts between shots in one clip and upscales to 3840×2176. It ran in ComfyUI on day one.
MiniMax opens H3's weights
MiniMax has published the weights of H3, the AI video model it launched on 31 July: a 33-billion-parameter transformer that generates picture and stereo sound together. The licence leaves out the EU, UK, South Korea and US.
Grok Imagine 1.5: seven references
xAI's Grok Imagine Video 1.5 now does text-to-video and reference-to-video, with up to seven image and voice references per AI video and native 1080p. In the Grok apps and, partly, in the xAI API.