Nico Ranieri

RokRol Builds the Next Short-Video Paradigm with AI-Powered Collaborative Storytelling RokRol has opened a private beta of its AI-powered short-video platform - designed not to optimize individual clips for virality, but to redefine how short-form video is created in the first place. RokRol introduces a new creative paradigm for short video: collaborative storytelling. Instead of isolated posts and passive consumption, creators on RokRol build stories together—layer by layer—by weaving multiple short videos into a shared narrative. From Clips to Stories: Rethinking Short Video Most short-video platforms today are built around a single unit of value: the standalone clip. Algorithms reward speed, frequency, and individual performance, often pushing creators into repetitive formats and endless scrolling. RokRol takes a fundamentally different approach. At the core of the platform is a video relay model, where creators can contribute sequential videos to an evolving storyline. Each video extends the narrative rather than competing for attention. AI assists throughout the process—supporting video generation, remixing, and creative iteration—while creators remain fully in control of the final result. The outcome is a feed that feels less like an algorithmic stream and more like a living story world. A New Creative Model for the AI Era As generative AI rapidly lowers the cost of video creation, the challenge is no longer how to generate content—but how to make it meaningful. RokRol positions AI not as an autonomous creator, but as a creative co-pilot: a tool that accelerates production while preserving human intent, authorship, and originality. Creators on RokRol can: Generate and remix short videos with AI assistance Join ongoing video relays and extend existing stories Collaborate with others to build narratives instead of posting isolated clips “We believe the next phase of short video isn’t just better generation—it’s a better creative structure,” said the RokRol team. “RokRol is built on the idea that stories become more powerful when they’re created together.” Early Traction and Creator Incentives RokRol is currently invite-only during its private beta. Early participants receive 10× points incentives, rewarding creators who help shape the platform and its emerging storytelling formats in this foundational stage. What Comes Next Looking ahead, RokRol plans to expand its AI-assisted creative toolkit and introduce richer narrative mechanics for relays and story chains—further strengthening collaborative creation as a first-class primitive. A broader public launch is planned following the beta period. RokRol’s Vision Short video doesn’t need to be faster. It needs to be deeper, more social, and more creative by design. RokRol isn’t just another short-video app. It’s an attempt to build the next short-video paradigm.

3

Unreal Engine 6 will be just a single neural network with 44x faster ray tracing? Or maybe OpenAI will be faster 😏 Last week, researchers from Microsoft (quite quietly) released the most interesting paper this year for me. They trained a full end-to-end mesh rendering pipeline using transformers. The idea is relatively straightforward. We describe the scene as a series of triangles with material properties and feed it as tokens into transformers (just as you would feed text into GPT). On scenes with up to 4k triangles and 512x512 res, it had ~44x speedup compared to a "classical" ray tracing rendering engine. And it works surprisingly well. We have to take into account this is very early research, but imagine if we trained this end-to-end on real-life data (e.g. games). We already use super-resolution a lot, but this would be a whole new level. Btw, reproducing the whole experiment would only cost $2k–$3k (assuming 8x A100 40GB, 8 days of training). Super excited for the weekend to play more with it. Btw. It kind of scares me that we can do so much with relatively simple solutions. Eg. assuming it would scale as good as GPT for text, OpenAI having expertise and infrastructure could create a SOTA rendering engine out of nowhere just by buying data.

14

4 Kommentare

Why existing music source separation models are already so good (easily >10 dB SDR), but mixing engineers rarely use them? Our answer: They need music source **restoration**. We introduce the music source restoration (MSR) task, which not only aims to separate the instruments, but also restore them to a state before any EQ, Compression, Reverb, Distortion and codec artifacts (e.g. MP3). We introduce the first dataset for this task, and found that the restoration task is much harder than separation. Paper: https://lnkd.in/ggDs3dad Model + Code: https://lnkd.in/gG3Psqqq Dataset: https://lnkd.in/gwUu6w3C HuggingFace Spaces: https://lnkd.in/gs3hJsmf Work done with Zheqi Dai, Mark Plumbley and Qiuqiang Kong. We are organizing a MSR challenge - if you are interested in participating, co-organizing or just learning more about this, please reach out to myself (zyy0116@gmail.com) or Prof. Qiuqiang Kong (qqkong@ee.cuhk.edu.hk)!

271

16 Kommentare

Turning real-world data into usable Unreal Engine environments is still far harder than it should be. Big thanks to 80.lv for publishing a proper technical deep dive on SynTerra 2.0 — not a press release, but an honest look at the real workflow constraints teams face when bringing actual geography into UE5. The article breaks down why many “real-world” pipelines fail in production — streaming dependencies, GIS overhead, slow iteration — and how we approach the problem differently by generating real locations directly from satellite and aerial data, offline and inside Unreal Engine. If you’ve ever tried to build large-scale, real environments for simulation, automotive, digital twins, or serious real-time projects, you’ll immediately recognize the pain points discussed there. 👉 Read the full article: https://lnkd.in/gF7Y9R4a This space doesn’t need louder demos. It needs tools that actually survive real production. 80 Level Epic Games Epic Games Unreal Engine #UnrealEngine #UE5 #EpicGames #GameDev #Simulation #DigitalTwins #Geospatial #SatelliteData #RealTime3D #SynTerra

21

Sonoma Alpha has arrived on Poe! This highly intelligent, stealth model from a frontier US lab features a 2M token context window, image input, parallel tool support, and comes in reasoning (Sky) and non-reasoning (Dusk) versions. Sonoma Alpha is free to use during this testing period, with anonymized prompts and outputs logged for the developer. You can try Sonoma Sky at https://lnkd.in/g-dV2PdT, and Dusk at https://lnkd.in/gCCSzG7q, across all platforms, and via the Poe API.

16

🎨 𝐍𝐨𝐢𝐬𝐲 𝐈𝐬𝐨𝐥𝐢𝐧𝐞𝐬 𝐰𝐢𝐭𝐡 𝐒𝐜𝐞𝐧𝐞 𝐍𝐨𝐝𝐞𝐬 𝐢𝐧 𝐂𝐢𝐧𝐞𝐦𝐚 𝟒𝐃 If you’ve explored Dominik Ruckli’s “Druckli Tool,” then you’ve seen how cleverly he uses Scene Nodes, especially his method for generating clean isolines using 𝐆𝐫𝐚𝐝𝐢𝐞𝐧𝐭 𝐈𝐧𝐭𝐞𝐫𝐬𝐞𝐜𝐭𝐢𝐨𝐧. In our latest tutorial, we’re taking a 𝐦𝐨𝐫𝐞 𝐚𝐫𝐭𝐢𝐬𝐭𝐢𝐜 𝐚𝐧𝐝 𝐥𝐞𝐬𝐬 𝐭𝐞𝐜𝐡𝐧𝐢𝐜𝐚𝐥 𝐚𝐩𝐩𝐫𝐨𝐚𝐜𝐡 to isolines in Cinema 4D. Instead of relying on complex node math, we build a setup that’s: ✨ Easy to follow ✨ Flexible for experimentation ✨ Great for motion design and abstract animation The result is 𝐛𝐞𝐚𝐮𝐭𝐢𝐟𝐮𝐥 𝐧𝐨𝐢𝐬𝐲 𝐢𝐬𝐨𝐥𝐢𝐧𝐞𝐬 that mimic the look of the original method while giving you more creative freedom. 🎥 Watch the full tutorial here: https://lnkd.in/gsk9qHZ9 ________ iRender - We focus on people and the joy of creation. 🌐 irendering.net 🎬 https://lnkd.in/daJn99sz #irender #renderfarm #Cinema4D #SceneNodes #MotionDesign #3DArt #C4DTutorial #Isolines #ProceduralDesign #DigitalArt

1

🌐 𝗛𝗢𝗟𝗢𝗚𝗥𝗔𝗣𝗛𝗜𝗖 𝗧𝗔𝗕𝗟𝗘 𝗣𝗥𝗢𝗝𝗘𝗖𝗧𝗜𝗢𝗡 → 𝗪𝗵𝗲𝗿𝗲 𝗦𝗰𝗶-𝗙𝗶 𝗺𝗲𝗲𝘁𝘀 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗘𝘅𝗰𝗲𝗹𝗹𝗲𝗻𝗰𝗲 ✍️ Article by Ulrich Buckenlei for XR Stager Newsroom What if the boundary between digital vision and physical design disappeared? In his latest article, Ulrich Buckenlei explores how AI, XR, and holography merge to create interactive 3D tables that bring imagination to life. 💡 𝗞𝗲𝘆 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀: ✨ 𝗔𝗜 𝗮𝘀 𝗮 𝗗𝗲𝘀𝗶𝗴𝗻 𝗣𝗮𝗿𝘁𝗻𝗲𝗿 → Turning data into living visual systems 🔹 𝗛𝗼𝗹𝗼𝗴𝗿𝗮𝗽𝗵𝘆 𝗮𝘀 𝗮 𝗠𝗲𝗱𝗶𝘂𝗺 𝗼𝗳 𝗖𝗿𝗲𝗮𝘁𝗶𝗼𝗻 → Projecting 3D content into shared physical spaces 🧩 𝗖𝗿𝗲𝗮𝘁𝗶𝘃𝗲 𝗖𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻 → Where designers, engineers, and AI co-create in real time 👉 Read the full article: https://lnkd.in/ek4Y-T5s #XR #AI #Holography #SpatialDesign #Innovation #DigitalTransformation #UlrichBuckenlei #XRStager

15

7 Kommentare

🚀 Advanced Pose-Conditioned Motion Transfer Pipeline (ComfyUI + Wan 2.2 14B) I've been rigorously testing this ComfyUI workflow for a while, validating motion fidelity, memory behavior, and temporal coherence — specifically for film-grade AI video inference and VFX-ready motion synthesis. 🏆 Motion replication benchmark (test conclusion) SCAIL Pose is unequivocally the best model for replicating character motion from reference video. Unlike alternatives such as Wan Animate or SteadyDancer, SCAIL does not stretch or deform the main character to force-match reference height/width, making motion transfer structurally stable and visually reliable. During tests: Even SteadyDancer fails to reproduce motion with this level of spatial accuracy Wan Animate applies destructive stretching to enforce fit SCAIL preserves proportions and pose integrity while transferring motion No character distortion, no body stretch, no forced canvas fit Most precise skeletal motion alignment achieved so far Not even SteadyDancer can replicate motions this precise. The spatial motion accuracy delivered by SCAIL is currently unmatched. 🎯 Why this is critical for film and VFX This system unlocks high-value use cases: 🎬 Pre-Production Storyboard → video animatic generation with real pose conditioning Early blocking and motion tests without proportion corruption Fast iteration for camera + character choreography alignment 🎥 Production Use real actor or 3D pose passes as a performance steering layer Generate style-remixed motion layers (robotic, cinematic, expressive) Keep movement predictable, non-destructive, and re-directable 🧩 Post-Production / VFX Produce VFX-ready motion passes for compositing Recover missing frames with latent diffusion sampling Maintain character continuity across shots Create clean, stretch-free motion transfer for 7–10 sec or longer sequences 🧠 Memory-Efficient Large Model Inference This workflow integrates: Attention block swap control to reduce VRAM peaks Empty CLIP Vision initialization prior to pose injection SCAIL Pose embedding injection to steer Wan 2.2 14B diffusion sampling WAN VAE decode + VHS Video Helper Suite muxing H.264 MP4 export at 24 FPS, CRF 19, YUV420p 🔧 Technical pipeline overview CLIP Vision init (empty embeds) → VitPose detection → DW Pose conversion → SCAIL Pose embedding injection → WanSampler v2 latent diffusion sampling → WAN VAE decode → Frame + audio mux via VHS Video Helper Suite → H.264 MP4 export at 24 FPS resource: https://lnkd.in/dQMhFTfN

25

Big news: Meshcapade is now part of Epic Games! 🎮 The US-based game developer and creator of Unreal Engine is acquiring the Max Planck spin-off and will establish a presence in Tübingen’s Cyber Valley. The Meshcapade team will join Epic’s AI Research organization, contributing to technologies for Unreal Engine and MetaHuman. Founded in 2018, Meshcapade develops solutions for the creation and animation of digital humans. Its technology is based on the SMPL (Skinned Multi-Person Linear) body model, trained on thousands of 3D body scans and now the de facto standard for 3D human representation across academia and industry. By combining machine learning and advanced graphics, Meshcapade enables highly expressive 3D digital humans, capturing realistic motion directly from standard video input and lowering barriers for creators across gaming, film, and virtual production. “This is truly remarkable,” said Max Planck President Patrick Cramer. “It shows how quickly insights from basic research can find their way into practical applications – with global impact.” Congratulations to Naureen Mahmood, Talha Zaman, Michael J. Black, and the entire Meshcapade team on this milestone. 👏 Cc: Lina Chong Read more here: https://lnkd.in/dM4-S7vu

30

PrismLabs reports a 340% improvement in answer accuracy on multi-hop queries and a 65% reduction in hallucination rates by moving from a "naive RAG" system to a hybrid GraphRAG architecture. The main problem with the "naive RAG" was that the vector search retrieved semantically similar chunks, but wasn't able traverse the logical chain from cause to effect across document boundaries. Well, although the hybrid GraphRAG can assure causality and usually improves performance, I would challenge the numbers. 340% improvement in answer accuracy is a lot. The issue with the like-for-like comparison is that the architectures require different ways to organise information in the AI system. A system with well-organised information in the "naive RAG" architecture will outperform a system with poorly organised information in a GraphRAG system. Unfortunately, the vast majority of GraphRAG systems rely on LLMs to identify the graph elements and to build the graph. And benchmarks demonstrated that such solutions achieve even worse results than the "naive RAG" based ones. 340% improvement in answer accuracy tells a different story. Perhaps the graph was well designed, but the "naive RAG" was not. In any case, I would always strongly suggest that, if you have a thorough understanding of machine learning with graphs, go with the graphs. Otherwise, stick to simpler architectures. https://lnkd.in/dMAa3bs2

gamescom will once again be an important meeting place for the XR scene in 2025. The Meta-sponsored Global XR/AI Hack will take place immediately before the trade fair. In cooperation with the DLR Projektträger and the EDFVR - ERSTER DEUTSCHER FACHVERBAND FÜR VIRTUAL REALITY there will also be a get-together directly at gamescom, both organised by Thomas Riedel https://lnkd.in/dBv-azGB

22

1 Kommentar

Saw that "The Witcher 4" UE5 tech demo with the voxel trees for fast rendering at a distance? I figured... that should be easy to do in TinyBVH. :) So now TinyBVH can do voxel meshes! Attached video: CPU-only, switching between BVH and voxels. The ray tracing is a bit slow here because the leafs have alpha textures. The voxel representation is faster here. This data format is particularly suitable for GPU however, so as soon that is working, this should fly. Code in tiny_bvh_foliage.cpp in the dev branch of TinyBVH on Github: https://lnkd.in/e64fiFum

272

3 Kommentare

Most image problems don’t come from tools. They come from not understanding pixels. Ganz just released Pixel Anatomy on HeyGanz.com, a subscription-based learning platform for first-principles Compositing training. 54 minutes · 46 short videos This course explains how digital images are actually constructed: - What a pixel is - How resolution and scaling really work - RGB values, bit depth, and precision - Alpha channels and premultiplication No intuition. No guesswork. Just the data images are made of. #AlphaChromatica #TICA #HeyGanz

8

🧐 How can Dynamic Gaussian Splats be used in a Virtual Production Workflow to help actors interact with virtual props? In this blog the CoSTAR National Lab worked with Disguise to discover the potential of Dynamic Gaussian Splat rendering techniques in the production pipeline, including virtual production, virtual art department, and props.    Read about how they got on and watch the full "Office Daydream" film here:  https://lnkd.in/eYeu8Nx8   This is the third of three research projects in a series about Gaussian Splatting, an emerging capture and rendering technique which offers exciting opportunities for accessible, high-fidelity 3D representations.    To find out more, read our Substack blog introduction to the Gaussian Splat series: https://bit.ly/DGSIntro   ☕ Sign up to the CoSTAR National Lab's new Substack blog: https://lnkd.in/eUT_aEVi 💫 Find out more about CoSTAR Network: https://lnkd.in/eBJhtBFW 📢 Sign up to the CoSTAR Network newsletter for call announcements and more: https://lnkd.in/euuhAFWs Royal Holloway, University of London | Abertay University | National Film & Television School | University of Surrey | Disguise | BT Group | Pinewood Group Limited   Peter Richardson | Miles Bernie | Rebecca Gregory-Clarke | Cristen Caine | Johnny Johnson | Gemma Campbell #CoSTARNetwork #CoSTARNationalLab #CreativeTechnology #Prototyping #GaussianSplats #CreativeIndustries #Innovation #Immersive #RandD #Funding #AccessProgramme #Support #VirtualProduction #Enterprise #Access #LivePerformance #ScreenIndustries #Performance #MotionCapture, #AI, #SpatialAudio

7

1 Kommentar

If you get a chance to open up Boris FX Silhouette and import 48 frames of Rotosplines from Machine Learning and give me some insight on how they work for you, I would be super grateful. 🙏 GitHub is an easy way to grab the stuff as a ZIP archive, you dont need to be a wizard with coding. https://lnkd.in/gNHcre6Z

10

💡PractiLight: Practical Light Control Using Foundational Diffusion Models 🔗 https://lnkd.in/eWxpkHGw tl;dr: Diffusion models already “know” a lot about light transport. No need for massive finetuning that hurts generalization. The trick to control light? Knowing where (layers) and when (timesteps) to add guidance, and how to preserve identity.  💡 We found that light transport “lives” in self-attention layers. 💡💡For timesteps, you should start early, but not too early, and stop midway. Not so shocking: many-to-many interactions in self-attention mirror how light bounces and propagates in a scene. ➡️ Self-attention acts as an inductive bias for modeling light transport. ⚙️ Solution: a lightweight LoRA adapter that surgically taps into the right layers/timesteps. Trained on only 4,000 synthetic image pairs — yet capable of unlocking rich, plausible relighting. Our adapter guides generation while preserving the identity & structure of the input. ✔️ Strong generalization across domains ✔️ Less catastrophic forgetting ✔️ Plausible lighting everywhere ✔️ Super lightweight ✨ Bottom line: If you want practical, domain-agnostic light control, this is for you ! 📄 PractiLight: Practical Light Control Using Foundational Diffusion Models 🔗 https://lnkd.in/eWxpkHGw 👥 Yotam Erel · Rishabh Dabral · Vlad Golyanik · Amit Bermano · Christian Theobalt

55

1 Kommentar

What if we could create avatars from the most abundant source of human capture — YouTube videos? University of Tübingen, Imperial College London and KAUST (King Abdullah University of Science and Technology) present “AHOY! Animatable Humans under Occlusion from YouTube Videos with Gaussian Splatting and Video Diffusion Priors ” This work present AHOY, a method for reconstructing complete, animatable 3D Gaussian avatars from in-the-wild monocular video despite heavy occlusion. Existing methods assume unoccluded input—a fully visible subject, often in a canonical pose—excluding the vast majority of real-world footage where people are routinely occluded by furniture, objects, or other people. Reconstructing from such footage poses fundamental challenges: large body regions may never be observed, and multi-view supervision per pose is unavailable. They address these challenges with four contributions: (i) a hallucination-as-supervision pipeline that uses identity-finetuned diffusion models to generate dense supervision for previously unobserved body regions; (ii) a two-stage canonical-to-pose-dependent architecture that bootstraps from sparse observations to full pose-dependent Gaussian maps; (iii) a map-pose/LBS-pose decoupling that absorbs multi-view inconsistencies from the generated data; (iv) a head/body split supervision strategy that preserves facial identity. We evaluate on YouTube videos and on multi-view capture data with significant occlusion and demonstrate state-of-the-art reconstruction quality. We also demonstrate that the resulting avatars are robust enough to be animated with novel poses and composited into 3DGS scenes captured using cell-phone video. Checkout the links in the comments for more info on the project and the team behind it! #machinelearning #computervision #multimodalai #audiovisual #egocentricvision #deeplearning #sceneunderstanding #robotics

36

1 Kommentar