Quiet week on the frontier video beat (June 19-23, 2026): no new model or version from the big labs. The freshest dated item is an open-weights cluster from Meituan's LongCat team. They open-sourced [WBench](https://github.com/meituan-longcat/WBench), a multi-turn benchmark for interactive video world models ([project site](https://meituan-longcat.github.io/WBench/), [arXiv](https://arxiv.org/html/2605.25874v1)): 289 test cases, 1,058 interaction rounds, 22 world models across 5 dimensions. New leaderboard entries this window: LingBot-World (fast) on June 17 and DreamX-World (5B AR) on June 18. The same team also dropped [LongCat-AudioDiT](https://news.aibase.com/news/26786), an open-source zero-shot voice-cloning TTS model that works in waveform latent space. Standing snapshot: the [Artificial Analysis Video Arena](https://artificialanalysis.ai/video/arena) shows [text-to-video](https://artificialanalysis.ai/video/leaderboard/text-to-video) led by HappyHorse-1.0 and [image-to-video](https://artificialanalysis.ai/video/leaderboard/image-to-video) led by Dreamina Seedance 2.0, with a competing arena putting Kling v3 on top. Boards disagree; treat rankings and prices as a monthly reshuffle. Main topic: style consistency, also called look-locking, making the overall LOOK match across shots (palette, lighting, grain, lens character, art direction) so a sequence reads as one world. This is distinct from character consistency, a [prior episode](https://kling.ai/blog/kling-ai-3-0-multi-reference-inpainting-guide). Style drifts because each generation is [stateless](https://getimg.ai/guides/guide-to-seed-parameter-in-stable-diffusion), re-deriving a look from scratch. Techniques covered: - Style-reference inputs: Midjourney's [style-reference flag and codes](https://docs.midjourney.com/hc/en-us/articles/32180011136653-Style-Reference), style weight, and the [V7 update](https://updates.midjourney.com/style-references-for-v7/); plus [Veo Ingredients](https://www.veo3ai.io/blog/veo-3-image-reference-workflow-2026), Runway Subject-Scene-Style, and [Kling multi-reference](https://kling.ai/blog/kling-ai-3-0-multi-reference-inpainting-guide). - A frozen [style bible](https://www.jointellos.com/blog/create-fashion-lookbook-ai-step-by-step-guide) prompt block, consistent seeds as a tie-breaker, minting all keyframes in one image model, and a [moodboard/lookbook](https://www.mindstudio.ai/blog/build-ai-fashion-lookbook-generator-mindstudio). - The grade as equalizer: [LUTs in DaVinci Resolve](https://borisfx.com/blog/how-to-add-luts-to-davinci-resolve-in-4-steps/), [Shot Match](https://larryjordan.com/articles/how-to-color-match-clips-in-davinci-resolve-20/), white balance on [Offset](https://gamut.io/fixing-white-balance-with-offset-and-gamma-in-davinci-resolve/), and [color-matching cameras](https://pixflow.net/blog/color-match-cameras-davinci-resolve/). - Power-user escape hatch: [IPAdapter](https://github.com/cubiq/ComfyUI_IPAdapter_plus) [style transfer](https://www.runcomfy.com/comfyui-workflows/comfyui-ipadapter-plus-style-transfer-made-easy). Bench on your own shots at the [Video Arena](https://artificialanalysis.ai/video/arena). Next up: the assembly edit in DaVinci Resolve, then color and grade.