VAST (Project Eden) has released EdenLive, a generative renderer that transforms a rough proxy stream from a 2D/3D scene, game engine, or standard recording into finished video in real time. World state and logic remain on the code and engine side, while the neural network is responsible only for the visuals. Weights of approximately 210 GB are available on Hugging Face, the code is published on GitHub, and the smaller 5B model, according to the authors' measurements, outputs 720p at a rate of approximately 24.7 FPS on a single consumer RTX 5090 card.

image
image
image

What Happened

The VAST team, as part of the Eden project, released EdenLive, a real-time generative renderer with open weights and code. Control is built on a post-trained DINOv3 encoder: it encodes proxy stream frames and feeds features into a diffusion transformer, which renders the final image. Two branches are offered: Wan2.1-I2V-14B at 480p resolution and Wan2.2-TI2V-5B at 720p resolution; each operates in bidirectional and few-step streaming modes with chunks of 4 steps, while the TAEHV decoder ensures more than 24 FPS. The 5-billion parameter version on an RTX 5090 card via the LightX2V accelerator achieves approximately 24.7 FPS at a resolution of 1280x704, and the repository contains demo reels several minutes long. The main code and DiT/VAE weights are published under the Apache 2.0 license; the flux-c2f component based on FLUX.2 [klein] 9B is required only for generating the first frame and is distributed under non-commercial terms. The project does not have a ready-made cloud API—the service currently operates only in self-hosted mode.

Context

The EdenLive concept is built on a division of labor between the engine and the model: the evolution of world state is calculated by standard code or a game engine, while the neural network is responsible only for visual synthesis. This removes the task of learning state transitions from the diffusion model, allowing the entire parameter budget limited for real-time use to be dedicated to image quality. The source of control is not a specialized control branch, but a universal post-hoc DINOv3 encoder, making the conditionality transferable between different engines and sources—2D/3D scenes or recordings—without retraining the renderer. The release itself is assembled from ready-made open components: base Wan models, the DINOv3 encoder, the TAEHV decoder, the LightX2V accelerator, and FLUX.2 [klein] for the first frame, making it more of a strong engineering composition than a new scientific paradigm. This hybrid contrasts with end-to-end action-conditioned world models, which in their pure form lose precise signals like health bars, cooldowns, and NPC states: in the EdenLive scheme, the engine still calculates all of this, while the model merely redraws the image.

Why This Matters for the Industry

For studios and media teams, a ready-made base for interactive previews is emerging: a rough scene in the engine immediately looks like a final render without waiting for art assets to be ready. Open 5B and 14B weights under Apache 2.0 allow embedding the pipeline into proprietary products, fine-tuning it for specific styles, and avoiding vendor lock-in. The cost of the final image drops to the level of a single consumer GPU, meaning rendering and rough art risk becoming commodities, while defensible value shifts to where world state lives—in engine code, logic, and content. From the decoupling of "simulation plus visuals," products can already be built: interactive game previews before art completion, v2v video stylization, and generative rendering over existing scenes.

Why This Matters for Users

Downloading approximately 210 GB of weights from Hugging Face and deploying the code from GitHub is possible today: the smaller 5B model operates at 720p on a single RTX 5090-class card, and minute-long demo reels in the repository show the results before installation. The practical value at the start lies in experiments, video stylization, and internal previews using a "prompt, scene, render" pipeline. Full real-time mode requires Linux, an NVIDIA GPU with 80 GB of video memory, and CUDA 13, so the barrier to entry is significantly higher than the phrasing about 720p on a single consumer card suggests. There is no cloud API—everything is calculated locally on your own hardware, and the workbench with FLUX first-frame generation is permitted exclusively for non-commercial projects.

What Is Currently Unknown / Limitations

All available quality assessments are demo reels and FPS measurements by the authors themselves; there are no quantitative metrics for temporal consistency, control accuracy, or drift in long streams in the release materials. Performance figures refer to different configurations: the shown approximately 24.7 FPS at 720p was achieved on an RTX 5090 via LightX2V, while the full real-time pipeline requires a card with 80 GB of video memory and CUDA 13, meaning the actual barrier to entry is likely higher than the headline promises. Independent benchmarks on other hardware and reports on the stability of long sessions have not yet been published, and the non-commercial license for the first-frame block limits direct commercial packaging.

Sources

Author

Look at AI, editorial team