nvidia/ChronoEdit-14B-Diffusers
171149
1---2license: apache-2.03language:4- en5- zh6pipeline_tag: other7tags:8- image editing9- video generation10library_name: diffusers11---12 13<p align="center">14 <img src="https://cdn-uploads.huggingface.co/production/uploads/658529d61c461dfe88afe8e8/tto-qt1BmKdhObReEKBq2.png" width="400"/>15<p>16 17<p align="center">18 💚 <a href="https://research.nvidia.com/labs/toronto-ai/chronoedit/"><b>ChronoEdit</b></a>    ｜    🖥️ <a href="https://github.com/nv-tlabs/ChronoEdit">GitHub</a>    |   🤗 <a href="https://huggingface.co/collections/nvidia/chronoedit">Hugging Face</a>   |   🤖 <a href="https://huggingface.co/spaces/nvidia/ChronoEdit">Gradio Demo</a>   |    📑 <a href="https://arxiv.org/abs/2510.04290">Paper</a>    19<br>20 21-----22# ChronoEdit: Towards Temporal Reasoning for Image Editing and World Simulation23 24### Description:25**ChronoEdit-14B** enables physics-aware image editing and action-conditioned world simulation through temporal reasoning. It distills priors from a 14B-parameter pretrained video generative model and separates inference into (i) a **video reasoning stage** for latent trajectory denoising, and (ii) an **in-context editing stage** for pruning trajectory tokens. ChronoEdit-14B was developed by NVIDIA as part of the **ChronoEdit** family of multimodal foundation models. 26This model is ready for commercial use.27 2829 30### License/Terms of Use31Governing Terms: Use of this model is governed by the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). Additional Information: [Apache License Version 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md).32 33Important Note: If you bypass, disable, reduce the efficacy of, or circumvent any technical limitation, safety guardrail or associated safety guardrail hyperparameter, encryption, security, digital rights management, or authentication mechanism contained in the Model, your rights under License Agreement will automatically terminate. (edited) 34 35### Deployment Geography36Global37 38 39### Use Case40Researchers and developers for:41- Physics-aware in-context image editing 42- Action-conditioned world simulation (PhysicalAI) 43- Benchmarking multimodal foundation models 44 45 46### Release Date47- Hugging Face: 10/29/2025 via [ChronoEdit](https://huggingface.co/collections/nvidia/chronoedit).48 49- GitHub: 29/10/2025 via [GitHub Repo Link](https://github.com/nv-tlabs/ChronoEdit)50 51 52## References(s)53- ChronoEdit: Temporal Reasoning for In-Context Image Editing (Preprint, 2025) 54- Related NVIDIA works: [Cosmos](https://www.nvidia.com/en-us/ai/cosmos/),[Gen3C](https://research.nvidia.com/labs/toronto-ai/GEN3C/),[DiffusionRenderer](https://research.nvidia.com/labs/toronto-ai/DiffusionRenderer/),[Difix3D](https://research.nvidia.com/labs/toronto-ai/difix3d/)55 56 57## Model Architecture58**Architecture Type:** Diffusion Transformer 59**Network Architecture:** Custom temporal denoising transformer 60**Base Model:** Pretrained video generative model (14B parameters) 61**Number of Parameters:** ~1.4 × 10^10 62 63 64---65 66 67## Input68**Input Type(s):** Image + Text (instruction) 69**Input Format:** 70- Text: UTF-8 string 71- Image: RGB (`.png`, `.jpg`) 72**Input Parameters:** 73- Image: Two-Dimensional (2D) Red, Green, Blue (RGB), variable resolution (recommended ≤1024×1024) 74- Text: One-Dimensional (1D), up to ~300 tokens 75**Other Properties Related to Input:**76- Image Resolution: 1280 x 720 or 720 x 1280 or 960 x 960 or 1024×102477 78 79 80 81## Output82**Output Type(s):** Image 83**Output Format:** RGB (`.png`) 84**Output Parameters:** Two-Dimensional (2D) Red, Green, Blue (RGB), resolution configurable 85**Other Properties Related to Output:**86- Image Resolution: 1280 x 720 or 720 x 1280 or 960 x 960 or 1024×102487 88 89Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions. 90 91 92---93 94 95## Software Integration96**Runtime Engine(s):** 97- PyTorch / Diffusers 98- Triton Inference Server (optional) 99 100 101**Supported Hardware Microarchitecture Compatibility:** 102- NVIDIA Ampere 103- NVIDIA Blackwell 104- NVIDIA Hopper 105- NVIDIA Lovelace 106 107 108**Preferred Operating Systems:** 109- Linux 110 111 112The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.113 114 115---116 117 118## Model Version(s)119- ChronoEdit-14B v1.0 (initial public release, 03/2025) 120 121 122---123 124 125## Training, Testing and Evaluation Datasets126 127 128## Training Dataset:129- Synthetic world interaction data (robot arm manipulation, object picking, temporal consistency) 130- Open-domain video-text corpora (research use only) 131 132 133**Data Modality:**134Image, Text, Video135 136 137**Image Training Data Size:**1381 Million to 1 Billion Images139 140 141**Text Training Data Size:**142Less than 10,000 Hours143 144 145**Data Collection Method:** Hybrid: Synthetic, Automated, Human146 147 148**Labeling Method:** Hybrid: Synthetic, Automated, Human149 150 151**Properties:**152- Modalities: 10 million image and text pair153- Nature of the content: Synthetic world interaction data (robot arm manipulation, object picking, temporal consistency) 154- Linguistic characteristics: Natural Language155 156 157---158 159 160## Testing Dataset161- Held-out portion of training dataset for action-conditioned tasks.162 163 164**Data Collection Method:**165Automated166 167 168**Labeling Method by dataset:**169Automated170 171 172**Properties:**173- Modalities: 500 million image and text pair174- Nature of the content: Robot world interaction data (robot arm manipulation, object picking, temporal consistency) 175- Linguistic characteristics: Natural Language176 177 178---179 180 181## Evaluation Dataset182- Held-out portion of training dataset for action-conditioned tasks.183 184 185**Data Collection Method:**186Automated187 188 189**Labeling Method by dataset:**190Automated191 192 193**Properties:**194- Modalities: 500 million image and text pair195- Nature of the content: Robot world interaction data (robot arm manipulation, object picking, temporal consistency) 196- Linguistic characteristics: Natural Language197 198 199---200 201 202## Inference203**Acceleration Engine:** TensorRT, Triton204**Test Hardware:** NVIDIA H100, NVIDIA B200205 206 207---208 209 210## Ethical Considerations:211NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. 212For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.213 214 215Users are responsible for model inputs and outputs. Users are responsible for ensuring safe integration of this model, including implementing guardrails as well as other safety mechanisms, prior to deployment.216 217 218Please make sure you have proper rights and permissions for all input image and video content; if image or video includes people, personal health information, or intellectual property, the image or video generated will not blur or maintain proportions of image subjects included.219 220 221Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).222 223 224### Plus Plus (++) Promise225 226 227We value you, the datasets, the diversity they represent, and what we have been entrusted with. This model and its associated data have been:228 229 230- Verified to comply with current applicable disclosure laws, regulations, and industry standards.231- Verified to comply with applicable privacy labeling requirements.232- Annotated to describe the collector/source (NVIDIA or a third-party).233- Characterized for technical limitations.234- Reviewed to ensure proper disclosure is accessible to, maintained for, and in compliance with NVIDIA data subjects and their requests.235- Reviewed before release.236- Tagged for known restrictions and potential safety implications.237 238 239### Bias240 241 242Field | Response243:---------------------------------------------------------------------------------------------------|:---------------244Participation considerations from adversely impacted groups [protected classes](https://www.senate.ca.gov/content/protected-classes) in model design and testing: | None245Measures taken to mitigate against unwanted bias: | None246 247 248### Explainability249 250Field | Response251:------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------252Intended Task/Domain: | Image Editing, Text Prompt253Model Type: | Transformer254Intended Users: | Physical AI developers.255Output: | Two-Dimensional (2D) Red, Green, Blue (RGB) Image256Describe how the model works: | We take an image as input, encode it using Cosmos tokenizer to latent space. We then use our model which is a transformer-like architecture to modify the image.257Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable.258Technical Limitations & Mitigation: | The proposed method relies only on synthetic data for Physical AI scenarios, which might limit the generalization ability if the target scenario is not in the pre-generated SDG dataset. 259Verified to have met prescribed NVIDIA quality standards: | Yes260Performance Metrics: | Qualitative and quantitative evaluation, including human and vision-language model (VLM) assessments on Action Fidelity, Identity Preservation and Visual Coherence Metrics.261Potential Known Risks: | This model is only trained on synthetic data generated for Physical AI use case. Testing on cases that are not related to Physical AI may yield unsatisfactory and unexpected results.262Licensing: | Governing Terms: Use of this model is governed by the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). Additional Information: [Apache License Version 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md).263 264 265### Privacy266 267 268Field | Response269:----------------------------------------------------------------------------------------------------------------------------------|:-----------------------------------------------270Generatable or reverse engineerable personal data? | No271Personal data used to create this model? | None Known272How often is dataset reviewed? | Before Release273Is there provenance for all datasets used in training? | Yes274Does data labeling (annotation, metadata) comply with privacy laws? | Yes275Is data compliant with data subject requests for data correction or removal, if such a request was made? | No, not possible with externally-sourced data.276Applicable Privacy Policy | https://www.nvidia.com/en-us/about-nvidia/privacy-policy/277 278 279### Safety280 281 282Field | Response283:---------------------------------------------------|:----------------------------------284Model Application Field(s): | World Generation285Describe the life critical impact (if present). | Not Applicable <br>286Use Case Restrictions: | Abide by [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). Additional Information: [Apache License Version 2.0](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md).287Model and dataset restrictions: | The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to.288 289 290## Citation291```bibtex292@article{wu2025chronoedit,293 title={ChronoEdit: Towards Temporal Reasoning for Image Editing and World Simulation},294 author={Wu, Jay Zhangjie and Ren, Xuanchi and Shen, Tianchang and Cao, Tianshi and He, Kai and Lu, Yifan and Gao, Ruiyuan and Xie, Enze and Lan, Shiyi and Alvarez, Jose M. and Gao, Jun and Fidler, Sanja and Wang, Zian and Ling, Huan},295 journal={arXiv preprint arXiv:2510.04290},296 year={2025}297}298 299```