Zing-0.5 - Loopit open-source real-time interactive video world model

What is Zing-0.5?

Zing-0.5 is an open-source real-time interactive video world model from Loopit, with 5-byte parameters and a Wan2.2-TI2V-5B base, licensed under the Apache 2.0 license. Its core capability is "joint control," where keyboard spatial manipulation and natural language semantic rewriting both enter Causal DiT. Users can move their character while simultaneously rewriting the world as it unfolds with text; new commands are applied to the current world rather than resetting the scene. The model learns to "modify while playing" by switching Prompt during training, actively perturbing historical frames to combat error accumulation. A single RTX 5090 GPU can generate over 24 FPS in real-time, with a streaming inference cost of approximately 6 cents per minute. It ranked first among real-time world models in the WBench benchmark jointly conducted by Meituan LongCat and Fudan University.

Main functions of Zing-0.5

• Joint controlKeyboard spatial operations (W/A/S/D for movement + I/J/K/L for view) work together with natural language semantic rewriting to change the world with text while controlling movement. New commands enter the current world rather than replacing the scene.
• Real-time streaming generationA single RTX 5090 card achieves over 24 FPS, with an inference cost of approximately 6 cents per minute, and supports continuous generation of infinitely long time sequences.
• Mid-process semantic rewritingText commands (such as "blizzard" or "dragon breathes fire") can be inserted at any time during the generation process, and the model retains the visual and operational history, allowing the rewriting to occur in the current world.
• Multiple objects responding simultaneouslyA single command can simultaneously drive changes to multiple different objects (e.g., a person kneeling in prayer + multiple statues opening their eyes and glowing).
• Long-range consistencyThe character's appearance, scene structure, and art style (such as pixel art) remain stable during continuous generation, without drifting or changing art style.
• Anti-error accumulationDuring training, historical footage is actively disturbed, and the model learns to "catch its own mistakes," preventing it from going further astray over a long period of time.
• motion control: 8-dimensional continuous motion signal (values [0,1]), supporting forward/backward/left/right movement and four-way viewpoint changes, encoded as frame-by-frame action residual.
• Local deployment5B parameters can run on a single card, Linux + PyTorch environment, 80GB+ VRAM or sliding attention configuration (RTX 4090 level is also possible), output H.264 MP4.
• Open source and freeReleased under the Apache 2.0 license, with code and weights available on GitHub, ModelScope, and HuggingFace, and trained on Wan2.2-TI2V-5B.

Technical Principles of Zing-0.5

• Base modelBased on Wan2.2-TI2V-5B training, with 5B parameters, the core generator network is Causal DiT (Causal Diffusion Transformer), which, together with causal KV cache, supports infinitely long time series autoregressive inference.
• Dual control link architectureTwo independent but shared control channels enter Causal DiT: spatial operations follow the "action chain," and semantic rewriting follows the "text chain." Together, they determine the content generated next.
• Action residualThe W/A/S/D motion and I/J/K/L viewpoints constitute an 8-dimensional continuous motion signal (values [0,1]), which is converted into a frame-by-frame action residual by sin/cos encoding and causal temporal convolution (less than 10M parameters). This residual is then injected and generated after being aligned frame-by-frame with the corresponding video token.
• Text chain (cross-attention)Natural language instructions are encoded into language tokens by a language model and written into the current generation through cross-attention, achieving semantic rewriting midway.
• Two-chain decoupling and continuous parallelismThe action chain is not interrupted when the text conditions change, so it can achieve effects such as "a dragon flying in the original direction while breathing fire" that are rewritten while being operated.
• Training means learning to "rewrite midway".During training, the Prompt is switched midway through autoregressive generation to preserve the generated visual and operational history. The subsequent text is predicted using new text conditions. The rewriting ability is a native ability developed during training, rather than an added feature added during inference.
• Anti-error accumulation trainingThe "history" generated by autoregression is the model's own output, and errors accumulate. Zing-0.5 actively perturbs some historical images during training (adding noise, slight blurring, color shift, and viewpoint changes) to simulate real errors, keeping the training target clean and forcing the model to learn to distinguish the real state from imperfect history.
• Pure training solution, zero inference overheadThe anti-accumulation capability is fully internalized in the weights, and online inference does not require an additional error correction model, thus not increasing latency or computing power.
• Four-step DMD distillation samplingCompressing diffusion sampling to 4 steps is key to single-card real-time (>24 FPS) and low cost (about 6 cents/minute).
• Sliding window attention-based video memory adaptation80GB+ of video memory --local-attn-size 97 --sink-size 9RTX 4090 level --local-attn-size 33 --sink-size 5It maintains a long context within limited GPU memory by using local attention and sink tokens.
• Output pipelineThe video frames are patch-embedded into tokens and fed into Causal DiT. After the predicted frames are decoded, they are output as H.264 encoded 24 FPS MP4.

How to use Zing-0.5

• Environmental preparationHardware and system preparation: Linux system + NVIDIA GPU (80GB+ VRAM recommended, such as H100/A100; RTX 4090 level can use sliding window attention configuration), basic environment installation: Python 3.11, PyTorch 2.9, CUDA.
• Get code and model
  • Clone code repository:git clone https://github.com/seedleap/zing-world-model.gitEnter the directory cd zing-world-model。
  • Install dependencies:pip install -r requirements.txt。
  • Download model weights:pip install modelscopeThen execute modelscope download --model seedleap/Zing-0.5 --local_dir ./Zing-0.5。
  • Confirm directory structure: weights include generator/model.pt(Generator) and pretrained/（text_encoder、tokenizer、vae）。
• Running inference
  • Text-driven mode (Action T2V): Execution run.sh, specify --pretrained-dir、--checkpoint、--messages examples/case3_action_t2v.jsonlOutput directory and random seed.
  • Image initialization mode (Action TI2V): Also uses... run.shReplace examples/case4_action_ti2v.jsonl Image generation and video configuration.
  • Select the attention parameter based on video memory: 80GB+ --local-attn-size 97 --sink-size 9Below 80GB --local-attn-size 33 --sink-size 5; Complete historical attention --local-attn-size -1 --sink-size 0。
• Interactive operation
  • Keyboard controls: W/A/S/D control forward, backward, left, and right movement; I/J/K/L control camera rotation (internal is an 8-dimensional continuous vector with values [0,1]).
  • Mid-process semantic rewriting: Input natural language commands (such as "blizzard" or "dragon's mouth breathes fire") during generation, and the world responds in real time without resetting the scene.
• View outputGenerates a 24 FPS MP4 file encoded in H.264 and saves it in the specified location. --output-dir Table of contents.
• Zero-deployment experience (optional)If you don't want to deploy locally, you can wait for the official product: the corresponding function is expected to be launched in the overseas version of the Loopit APP within two weeks, allowing you to experience the real-time interactive world directly in the app.

Zing-0.5's core advantages

• The raw strength of being number one on the real-time leaderboardThe WBench benchmark score is 81.0, ranking second overall, only behind a non-real-time bidirectional model, and first in the real-time world model ranking.
• "Play and modify at the same time" is in a league of its own.Most real-time world models only support "exploring the scene," while Zing-0.5 is one of the few models that can rewrite the world with text while manipulating movement. Joint control is the core differentiator.
• Rewrite but don't change the worldThe new instructions are precisely written into the current world, only changing what needs to be changed (adding blizzards, making dragons breathe fire), while leaving the rest (characters, art style, distant scenes) completely untouched. Similar models can often "incidentally replace the entire world".
• Long timings do not deviateDuring training, the model actively perturbs historical images to learn to correct errors, maintaining stability in character appearance and scene structure over a long period of time; moreover, it is a pure training phase solution, requiring no additional error correction model during inference and without increasing latency.
• Low cost and low barrier to entryWith 5B parameters and single-card deployment, a single RTX 5090 can achieve over 24 FPS, and one minute of streaming inference costs only about 6 cents, enabling real-time interaction with consumer-grade graphics cards.
• Lightweight and elegant architectureThe action control chain has less than 10M parameters (sin/cos encoding + causal convolution → frame-by-frame action residual). The dual control chains enter Causal DiT together, resulting in a simple and efficient design.
• True open sourceApache 2.0 license, with the entire set of code, authority, and deployment documentation open, making it commercially accessible. The Wan2.2-TI2V-5B base is mature and reliable.

Zing-0.5 project address

• Github repository：https://github.com/seedleap/zing-world-model
• HuggingFace model library：https://huggingface.co/seedleap/zing-0.5
• ModelScope：https://modelscope.cn/models/seedleap/Zing-0.5

Zing-0.5 Comparison with Similar Competitors

Application scenarios of Zing-0.5

• AI Interactive Entertainment / AI GamesPlayers use the keyboard to control their characters to explore a generated world, and can rewrite the plot and environment at any time with language (such as snow, summoning monsters), realizing a "generative game that everyone can play".
• "Playable Douyin" style content consumptionShort videos have transformed from "watching" to "playing," with each user creating different outcomes within the same content, thus reshaping the content platform landscape.
• UGC Interactive Content CreationCreators generate interactive video worlds for fans to experience, and viewers are no longer bystanders but participants, giving rise to entirely new content categories.
• Rapid Prototype ValidationGame teams can quickly generate interactive demos using world models without building engine scenes, thus validating gameplay and level ideas at low cost.
• Virtual live streaming / virtual anchorThe host drives the virtual world in real time, and viewers participate by rewriting scenes and events through bullet comments, enhancing the interactivity of the live stream.