Real‑Time 16B‑Parameter Nano Banana Model Open‑Sourced for Video Editing

In August the AI video race accelerated: ByteDance released Seedance 2.5, doubling single‑shot generation length, and MiniMax H3 opened at a third of the price of closed‑source flagship models. The common goal across the race is to generate longer, more coherent, and cheaper videos, but all solutions still require waiting for the full video to finish before any edits can be made.

On August 5 JD Open‑Source released JoyAI‑Video‑Edit, the first model that simultaneously satisfies three requirements—streaming architecture, real‑time speed, and usable quality. The model runs at 30 frames per second (FPS) on 720p video, delivering a stable end‑to‑end 24 FPS stream with unlimited length, allowing continuous on‑the‑fly visual effects.

Traditional video‑generation models are offline: they ingest an entire clip, process it, and output the result. Real‑time editing demands two hard constraints. The first is speed: video playback runs at 24–30 FPS, so any processing slower than that causes stutter. JoyAI‑Video‑Edit processes each incoming frame immediately and streams the edited frame out without waiting for future frames, analogous to simultaneous interpretation.

The second constraint is length. Existing streaming editors degrade after a few seconds to a couple of minutes because autoregressive diffusion accumulates errors, causing visual drift. To keep quality over long streams JoyAI‑Video‑Edit introduces a "bounded KV state inference" mechanism that limits the model’s memory to the most recent frames plus the first frame as a reference, keeping compute and memory constant regardless of video duration.

Architecturally, JoyAI‑Video‑Edit combines a multimodal large‑language‑model (MLLM) conditional encoder, a causal video VAE, and a 16‑billion‑parameter multimodal diffusion Transformer (DiT). The system is trained as an autoregressive diffusion editor. To overcome the prohibitive inference cost of a 16 B model, the team applied a training technique called SA‑DMD, which distills the diffusion process from many iterative steps down to just two steps, similar to training a professor to answer instantly after years of practice.

On a single Nvidia B200 GPU the latency breakdown is: VAE encoding 22 ms, DiT denoising 185 ms, VAE decoding 19 ms, yielding a total request‑to‑response latency of 226 ms and an end‑to‑end throughput of 30.1 FPS. Compared with prior streaming editors (StreamDiffusionV2, LiveEdit, SANA‑Streaming) that use 1.3–2 B parameters and are limited to 480p, JoyAI‑Video‑Edit’s 16 B model is 1.4–2× faster at 720p while delivering higher visual fidelity.

In the OpenVE‑Bench short‑video benchmark (the industry‑standard video‑editing suite) JoyAI‑Video‑Edit achieved a total score of 3.60, surpassing SANA‑Streaming (2.62), LiveEdit (2.00), XMax‑X2.0 (1.87) and StreamDiffusionV2 (1.23). It ranked first on four of five sub‑tasks, with a particularly strong advantage on local‑delete (4.06 pts), even exceeding all offline models. Its 3.60 score places it within the commercial offline model range (Runway Aleph 3.45, PixVerse 3.05).

On the LongV2VBench long‑video benchmark covering background replacement, global style, local addition, modification, and deletion, JoyAI‑Video‑Edit ranked first in every category with a total of 3.30, 1.59 points above the next best XMax‑X2.0. It also ran fastest at 30.19 FPS, 44.4 % faster than XMax‑X2.0 and more than twice the speed of SANA‑Streaming.

Human blind‑pairwise evaluations further confirm its superiority: JoyAI‑Video‑Edit was preferred over four competing streaming models with rates of 90 %, 87 %, 87 % and 81 %, and achieved a slight edge (48 % vs 44 %) against the strong offline model Bernini‑R.

The practical impact is a paradigm shift in video production. The "wait‑for‑render" loop disappears—adjusting color, style, or objects yields an almost instantaneous response (226 ms latency), turning post‑production into live interaction. This enables live‑streamed, on‑the‑fly variant generation (e.g., real‑time virtual try‑on), and introduces a programmable semantic layer between camera input and display, far beyond simple filters.

Beyond creative use‑cases, JD envisions leveraging real‑time editing for embodied AI. Training robots requires massive, high‑quality manipulation videos, which are scarce. JoyAI‑Video‑Edit can replace human hands with robot arms in video streams while preserving object positions, spatial relations, and motion trajectories, producing 30 FPS training data at scale. This capability fits into JD’s broader JoyAI model ecosystem (Vision‑Language Interaction, Talker, RA) aimed at building a "global physical‑world operation center" for logistics, live commerce, and autonomous systems.