Margiela @MargielaCSS

hyper developer Joined November 2024

Tweets

534
Followers

13K
Following

573
Likes

658

Margiela @MargielaCSS

2 days ago

@itsshara_ai wow, seen some incredible things being generated with SD2.5, need to try it out

2 0 1 70 0

View Details

Shara I Ai Video Creator @itsshara_ai

2 days ago

These shots are 100% AI-made. Created with Seedance 2.5. No reference images. No animators. Just prompts Prompt below:

171 76 953 43K 484

View Details

Alex Reinicke @Alex_Reinicke_

2 days ago

corgi cafe london at midnight. 24/7.

23 8 337 36K 54

View Details

Watcher.Guru @WatcherGuru

4 days ago

JUST IN: Elon Musk's SpaceX $SPCX surges over 8% back to $125

559 675 8K 728K 186

View Details

NEW MODEL: Roblox now has its own world model! DEFINITIONS: - World model: an AI model that learns to predict what happens next in an environment. - ControlNet-style conditioning: a technique from image generation where a side network feeds a precise spatial guide (here the draft frame and depth map) into the generator, forcing the output to follow that structure pixel for pixel. - Cross-attention: the mechanism that lets the generator "look up" side information (here, object descriptions) while producing each patch of the image, instead of having it baked into the input. - Latents: the compressed internal representation a generative model works in. - Self-forcing: a training trick that converts a video model built to generate whole clips offline into one that emits frames one at a time from its own previous output, which is what makes live streaming possible. - KV cache: the transformer's running memory of everything it has generated so far. Roblox Reality is a WIP "hybrid architecture" for photorealistic multiplayer games. The game engine, running in Roblox Cloud across 20–26+ edge data centers, holds authoritative world state (positions, velocities, materials, physics, collision, ray tracing) as the single source of truth and emits a low-quality draft frame plus a depth map, lighting/weather data, and structured object descriptions. A video world model called the Super Upsampler, running on H200/B200-class edge GPUs, takes that draft as conditioning and repaints each frame into 2K photorealistic video (target 60fps), streamed to the client. It is meant to replace the traditional setup where high-fidelity rendering demands powerful user hardware and a large art budget. Worth specifying imho: this is not a world model as we usually expect. It is not generating the world. The engine still renders a real draft frame with true physics and state, and the model only repaints pixels on top of it. Dense signals (draft frame + depth) inject via ControlNet-style nets, global lighting/weather via modulation, object descriptions via cross-attention. The name "Super Upsampler" is more honest imho. Roblox explicitly rejected "pure" video world models because they can't do multiplayer, rule enforcement, persistent state, or reliable control. Nothing is cached and inference is per-player, per-frame: appearance lives only as video latents regenerated continuously and never saved, and each player runs their own model generating their own view -> pretty impressive and compute intensive imho. Durable world state (position, velocity, materials, direction) lives server-side in the engine's data model and is shipped identically to every client. Appearance is ephemeral latents each client regenerates. Player input (WASD) feeds straight into the model as conditioning, with client-side predictive simulation for immediate responsiveness, and locally-rendered upsampled avatars optionally overlaid to keep foreground actions low-latency. With regards to compute, Roblox owns 20–26+ edge data centers and its own proprietary GPUs, deliberately avoiding cloud provider dependency. The latency ask is ~30ms/frame, Roblox closes it with "self-forcing" (converts the offline model into autoregressive, frame-by-frame streaming generation), smaller models, key-value cache compression (needed at all for high resolution), and edge GPUs co-located with the game engine. Roblox did three acqui-hires in this space: Morpheus AI (self-forcing + long context to fight drift), Lucid AI (the "Game Cartridge Harness" that wraps deterministic engine logic around the model so rules hold), and Dynamics Lab (real-time image-to-world generation).