Hugging Face's logo

nvidia
/
Qwen-Image-Flash

Text-to-Image
Diffusers
Safetensors
QwenImagePipeline
nvidia
ModelOpt
AutoModel
qwen-image
dmd2
few-step
Model card Files
xet
 Community
1
jingyux-nv commited on
Commit
4873922
·
verified ·
1 Parent(s): 20fe4b8

Update Qwen-Image-Flash model card

Browse files
Files changed (1) hide show
  1. README.md +278 -94
README.md CHANGED
@@ -1,36 +1,164 @@
1
  ---
2
- license: apache-2.0
3
- language:
4
- - en
5
- library_name: diffusers
6
  pipeline_tag: text-to-image
7
- base_model: Qwen/Qwen-Image
 
 
 
 
 
 
8
  tags:
 
 
 
9
  - qwen-image
10
  - dmd2
11
  - few-step
12
  - text-to-image
13
  ---
14
 
15
- # Qwen-Image-Flash
16
 
17
- Qwen-Image-Flash is an experimental **4-step DMD2-distilled** text-to-image model based on
18
- [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image). It keeps the original
19
- Qwen-Image tokenizer, Qwen2.5-VL text encoder, VAE, and transformer architecture,
20
- while replacing the transformer weights with the distilled student and configuring
21
- the scheduler for the student's four-step shift-3 trajectory.
22
 
23
- This checkpoint is designed for exactly four ODE/FlowMatch Euler transformer evaluations.
24
- All packaged learned weights are BF16. This is a full Diffusers pipeline, not a LoRA
25
- and not a quantized model.
26
 
27
- This is a derivative release and is not affiliated with or endorsed by the Qwen team
28
- or Alibaba unless separately stated by the publisher.
29
 
30
- ## Quick start
31
 
32
- The artifact was produced with `diffusers==0.38.0` and `transformers==5.5.0`.
33
- Install the tested library versions plus Accelerate and Safetensors:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34
 
35
  ```bash
36
  pip install "diffusers==0.38.0" "transformers==5.5.0" accelerate safetensors
@@ -64,84 +192,140 @@ image = pipe(
64
  image.save("qwen-image-flash.png")
65
  ```
66
 
67
- For lower accelerator-memory use, replace `.to("cuda")` with
68
- `pipe.enable_model_cpu_offload()`.
 
 
 
 
 
 
 
 
 
69
 
70
- ## Required inference settings
71
 
72
  - Use `num_inference_steps=4`.
73
- - Use `true_cfg_scale=1.0` and omit `negative_prompt`. The student learned from a
74
- teacher target with CFG 4.0,
75
- so inference-time CFG would apply guidance a second time.
76
- - Keep the packaged static shift-3 scheduler and do not pass custom `sigmas`. With four
77
- steps it produces effective sigmas `[1.0, 0.9, 0.75, 0.5, 0.0]`, matching training.
78
-
79
- ## Model details
80
-
81
- | Property | Value |
82
- |---|---|
83
- | Base model | `Qwen/Qwen-Image` |
84
- | Distillation | DMD2, multi-step |
85
- | Student checkpoint | `epoch_4_step_15999` (saved after 16,000 training steps) |
86
- | Sampling | 4-step deterministic ODE / FlowMatch Euler |
87
- | Trained schedule | `[1.0, 0.9, 0.75, 0.5, 0.0]` |
88
- | Teacher CFG target | `4.0` (internalized by the student) |
89
- | Transformer dtype | BF16 |
90
- | EMA | Disabled; this release uses the live student weights |
91
-
92
- The transformer contains the same 1,933 parameter tensors and architecture as the base
93
- Qwen-Image transformer. The tokenizer, text encoder, and VAE are copied unchanged from
94
- `Qwen/Qwen-Image` revision `75e0b4be04f60ec59a75f475837eced720f823b6`.
95
- The scheduler differs only in its deterministic four-step shift-3 configuration.
96
-
97
- ## Training data
98
-
99
- The training run consumed a 1,000,000-sample, 1024-by-1024 cache whose internal dataset
100
- identifier is `qwen-image-synthetic-dalle3-1m`. That identifier alone does not establish
101
- the generator, caption source, filtering/deduplication process, or usage rights. The
102
- publisher should document and review those details before public release. Distillation
103
- captions were English; inherited Chinese capability has not been evaluated for this student.
104
 
105
  ## Evaluation
106
 
107
- No quantitative benchmark or latency results are currently reported. Any evaluation should
108
- use `num_inference_steps=4` and `true_cfg_scale=1.0`.
109
-
110
- ## Limitations and responsible use
111
-
112
- - This model was distilled specifically for the four-step schedule above; other step counts
113
- or scheduler settings are out of distribution and may reduce quality.
114
- - Four denoising steps reduce transformer evaluations but do not reduce parameter count or
115
- memory requirements. The model is not quantized.
116
- - The model inherits limitations and potential biases from Qwen-Image and from its
117
- distillation data. Generated text may still be misspelled or spatially inconsistent.
118
- - The packaged pipeline has no safety checker. Deployers must add moderation and filtering
119
- appropriate to their use case.
120
- - Intended use is research and evaluation of text-to-image generation. Image editing,
121
- image understanding, high-stakes decisions, and generation of illegal or abusive content
122
- are out of scope.
123
- - Users are responsible for reviewing generated content and complying with applicable laws,
124
- licenses, privacy requirements, and the base model's usage guidance.
125
-
126
- ## Acknowledgements and citations
127
-
128
- This work is derived from [Qwen-Image](https://arxiv.org/abs/2508.02324) and uses
129
- [DMD2](https://arxiv.org/abs/2405.14867) for few-step distribution matching distillation.
130
-
131
- ```bibtex
132
- @misc{wu2025qwenimagetechnicalreport,
133
- title={Qwen-Image Technical Report},
134
- author={Chenfei Wu and Jiahao Li and Jingren Zhou and Junyang Lin and Kaiyuan Gao and Kun Yan and Sheng-ming Yin and Shuai Bai and Xiao Xu and Yilei Chen and Yuxiang Chen and Zecheng Tang and Zekai Zhang and Zhengyi Wang and An Yang and Bowen Yu and Chen Cheng and Dayiheng Liu and Deqing Li and Hang Zhang and Hao Meng and Hu Wei and Jingyuan Ni and Kai Chen and Kuan Cao and Liang Peng and Lin Qu and Minggang Wu and Peng Wang and Shuting Yu and Ting Yu and Wensen Feng and Xiaoxiao Xu and Yi Wang and Yichang Zhang and Yongqiang Zhu and Yujia Wu and Yuxuan Cai and Zenan Liu},
135
- year={2025},
136
- eprint={2508.02324},
137
- archivePrefix={arXiv}
138
- }
139
-
140
- @misc{yin2024improveddistributionmatchingdistillation,
141
- title={Improved Distribution Matching Distillation for Fast Image Synthesis},
142
- author={Yin, Tianwei and Gharbi, Micha\"el and Park, Taesung and Zhang, Richard and Shechtman, Eli and Durand, Fredo and Freeman, William T.},
143
- year={2024},
144
- eprint={2405.14867},
145
- archivePrefix={arXiv}
146
- }
147
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
 
 
 
 
2
  pipeline_tag: text-to-image
3
+ base_model:
4
+ - Qwen/Qwen-Image
5
+ license: other
6
+ license_name: nvidia-open-model-license
7
+ license_link:
8
+ https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license
9
+ library_name: Model Optimizer
10
  tags:
11
+ - nvidia
12
+ - ModelOpt
13
+ - AutoModel
14
  - qwen-image
15
  - dmd2
16
  - few-step
17
  - text-to-image
18
  ---
19
 
20
+ # Model Overview
21
 
22
+ ## Description:
 
 
 
 
23
 
24
+ The NVIDIA Qwen-Image-Flash model generates images from text prompts using a four-step, DMD2-distilled version of [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image). The model was distilled using [NVIDIA Model Optimizer](https://github.com/NVIDIA/Model-Optimizer) and [NVIDIA AutoModel](https://github.com/NVIDIA-NeMo/Automodel) while retaining the base model architecture. The packaged scheduler is configured for the four-step, shift-3 trajectory.
 
 
25
 
26
+ Qwen-Image-Flash is a full BF16 [Diffusers](https://github.com/huggingface/diffusers) pipeline. The checkpoint is designed for exactly four deterministic ODE/FlowMatch Euler transformer evaluations.
 
27
 
28
+ This model is ready for commercial or non-commercial use.
29
 
30
+ ### License/Terms of Use:
31
+
32
+ **Governing Terms:** Use of this model is governed by the NVIDIA Open Model Agreement([https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/])) .
33
+
34
+ **Additional Information:** [Apache License 2.0](https://choosealicense.com/licenses/apache-2.0/).
35
+
36
+ ## Deployment Geography:
37
+
38
+ Global
39
+
40
+ ## Use Case:
41
+
42
+ Developers and researchers evaluating few-step text-to-image generation, including creative-content prototyping and latency-sensitive image-generation workflows. Image editing, image understanding, and safety- or life-critical decision-making are outside the intended scope of this checkpoint.
43
+
44
+ ## Release Date:
45
+
46
+ Hugging Face 07/23/2026 via [nvidia/Qwen-Image-Flash](https://huggingface.co/nvidia/Qwen-Image-Flash)
47
+
48
+ # Reference(s):
49
+
50
+ - [Qwen-Image model card](https://huggingface.co/Qwen/Qwen-Image)
51
+ - [Qwen-Image Technical Report](https://arxiv.org/abs/2508.02324)
52
+ - [Improved Distribution Matching Distillation for Fast Image Synthesis (DMD2)](https://arxiv.org/abs/2405.14867)
53
+ - [NVIDIA Model Optimizer](https://github.com/NVIDIA/Model-Optimizer)
54
+ - [NVIDIA FastGen](https://github.com/NVlabs/FastGen)
55
+ - [NVIDIA NeMo AutoModel](https://github.com/NVIDIA-NeMo/Automodel)
56
+
57
+
58
+
59
+ ## Model Architecture:
60
+
61
+ **Architecture Type:** Diffusion Transformer (MMDiT)
62
+
63
+ **Network Architecture:** `QwenImageTransformer2DModel` in a `QwenImagePipeline`
64
+
65
+ **Base Model:** [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image)
66
+
67
+ **Number of Model Parameters:** 20.43B parameters in the denoising transformer; 28.85B learned parameters in the full pipeline when the text encoder and VAE are included.
68
+
69
+ The full pipeline contains a Qwen2.5-VL text encoder, Qwen tokenizer, 60-layer Qwen-Image transformer, Qwen-Image VAE, and FlowMatch Euler scheduler. The transformer has the same architecture as the base Qwen-Image transformer; its weights are replaced by the distilled student weights.
70
+
71
+ ## Input:
72
+
73
+ **Input Type(s):** Text
74
+
75
+ **Input Format(s):** String
76
+
77
+ **Input Parameters:** One-Dimensional (1D) sequence
78
+
79
+ **Other Properties Related to Input:** The release was distilled with English captions. Any inherited Chinese-language capability has not been evaluated for this model. The pipeline allows at most 1,024 prompt tokens and defaults to 512.
80
+
81
+ ## Output:
82
+
83
+ **Output Type(s):** Image
84
+
85
+ **Output Format:** Red, Green, Blue (RGB)
86
+
87
+ **Output Parameters:** Two-Dimensional (2D)
88
+
89
+ **Other Properties Related to Output:** The pipeline returns generated images, typically as Python Imaging Library (PIL) images. The tested output setting is 1024 × 1024 pixels. Use width and height values divisible by 16 to avoid automatic resizing.
90
+
91
+ Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA hardware and software frameworks such as CUDA and PyTorch, the model can achieve faster inference than CPU-only execution.
92
+
93
+ ## Software Integration:
94
+
95
+ **Supported Runtime Engine(s):**
96
+
97
+ - [Hugging Face Diffusers](https://github.com/huggingface/diffusers) 0.38.0
98
+ - [PyTorch](https://pytorch.org/)
99
+ - [Transformers](https://github.com/huggingface/transformers) 5.5.0
100
+
101
+ **Supported Hardware Microarchitecture Compatibility:**
102
+
103
+ - NVIDIA Hopper H100
104
+ - NVIDIA Blackwell B200
105
+
106
+ **Preferred Operating System(s):**
107
+
108
+ - Linux
109
+
110
+ The 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.
111
+
112
+ ## Model Version(s):
113
+
114
+ Qwen-Image-Flash was distilled with NVIDIA Model Optimizer 0.45.0.
115
+
116
+ ## Training and Evaluation Datasets:
117
+
118
+
119
+
120
+ ## Training Dataset:
121
+
122
+ **Source Prompt Dataset:** [ProGamerGov/synthetic-dataset-1m-dalle3-high-quality-captions](https://huggingface.co/datasets/ProGamerGov/synthetic-dataset-1m-dalle3-high-quality-captions), dataset is comprised of AI-generated images sourced from various websites and individuals, primarily focusing on [Dalle 3](https://openai.com/index/dall-e-3/) content.
123
+
124
+ **Image Generator:** [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image)
125
+
126
+ **Internal Training Dataset Identifier:** `qwen-image-synthetic-dalle3-1m`
127
+
128
+ **Data Modality:** Text, Image
129
+
130
+ **Text Training Data Size:** [Less than a Billion Tokens]
131
+
132
+ **Image Training Data Size:** [Less than a Million Images]
133
+
134
+ **Data Collection Method by dataset:** Hybrid: Automated, Synthetic
135
+
136
+ **Labeling Method by dataset:** Automated
137
+
138
+ **Properties:** Only text prompts from the ProGamerGov dataset were used; its original images were not used for training. Each prompt was passed to the original Qwen-Image model to generate a new synthetic image. The resulting dataset contains 1,000,000 English prompt–image pairs processed at 1024 × 1024, and a subset was used for Distribution Matching Distillation (DMD2) training.
139
+
140
+ ## Evaluation Dataset:
141
+
142
+ **Link:** MTBench, for more details, see [here](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge)
143
+ **Data Collection Method by dataset:** Hybrid: Manually-collected, Synthetic
144
+ **Labeling Method by dataset:** Hybrid: manually-labelled, Synthetic
145
+ **Properties:** 3,300 multi-turn dialogue sequences.
146
+
147
+ ## Inference:
148
+
149
+ **Acceleration Engine:** Hugging Face Diffusers with PyTorch
150
+
151
+ **Test Hardware:** NVIDIA Hopper H100
152
+
153
+ ## DMD2 Distillation
154
+
155
+ Qwen-Image-Flash was trained using multi-step DMD2 (Distribution Matching Distillation) to approximate the output distribution of Qwen-Image in four denoising steps. The student retains the base transformer's architecture and parameter count. The four-step design reduces the number of transformer evaluations but does not reduce the checkpoint's parameter count or memory requirements.
156
+
157
+ The teacher target used classifier-free guidance (CFG) 4.0 during distillation. That guidance is internalized by the student, so inference uses `true_cfg_scale=1.0` to avoid applying guidance a second time. The packaged static shift-3 FlowMatch Euler scheduler produces effective sigmas `[1.0, 0.9, 0.75, 0.5, 0.0]` over the required four steps.
158
+
159
+ ## Usage
160
+
161
+ Use Python 3.10 or newer and first install a CUDA-enabled PyTorch build compatible with the system's NVIDIA driver by following the [PyTorch installation instructions](https://pytorch.org/get-started/locally/). Then install the library versions used to produce and test the artifact, together with Accelerate and Safetensors:
162
 
163
  ```bash
164
  pip install "diffusers==0.38.0" "transformers==5.5.0" accelerate safetensors
 
192
  image.save("qwen-image-flash.png")
193
  ```
194
 
195
+ For lower accelerator-memory use, construct the pipeline without `.to("cuda")`, then enable model CPU offload:
196
+
197
+ ```python
198
+ pipe = QwenImagePipeline.from_pretrained(
199
+ model_id,
200
+ torch_dtype=torch.bfloat16,
201
+ )
202
+ pipe.enable_model_cpu_offload()
203
+ ```
204
+
205
+ Model CPU offload still requires enough accelerator memory for the approximately 40.86 GB BF16 transformer plus activations. `pipe.enable_sequential_cpu_offload()` lowers accelerator-memory use further at the cost of additional transfer overhead and slower inference.
206
 
207
+ **Required Inference Settings:**
208
 
209
  - Use `num_inference_steps=4`.
210
+ - Use `true_cfg_scale=1.0` and omit `negative_prompt`; the student has already internalized the teacher's CFG target.
211
+ - Keep the packaged static shift-3 scheduler and do not pass custom `sigmas`.
212
+
213
+
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
214
 
215
  ## Evaluation
216
 
217
+ We report **acceptance length (AL)** — the mean number of tokens emitted per decoding step (the accepted draft tokens plus the one bonus token from the target model). Higher is better; with draft length *d* the maximum is *d*+1.
218
+
219
+ Measurement setup: vLLM DFlash speculative decoding, greedy decoding (temperature 0), target `Qwen/Qwen-Image`, tensor-parallel 1 on **NVIDIA Hopper H100**. We list two draft lengths — **3** (for comparison with other published speculative-decoding cards) and **7** (DFlash `block_size` − 1, the recommended serving configuration).
220
+
221
+ Acceptance length on [MT-Bench](https://huggingface.co/spaces/lmsys/mt-bench) (80 prompts):
222
+
223
+
224
+ | Category | MT Bench Acceptance Rate |
225
+ | ------------------- | ------------------------ |
226
+ | writing | 2.93 |
227
+ | roleplay | 2.54 |
228
+ | reasoning | 3.09 |
229
+ | math | 4.54 |
230
+ | coding | 3.42 |
231
+ | extraction | 4.13 |
232
+ | stem | 2.77 |
233
+ | humanities | 2.43 |
234
+ | **Overall Average** | **3.23** |
235
+
236
+
237
+ Acceptance length on [SPEED-Bench](https://huggingface.co/datasets/nvidia/SPEED-Bench) (qualitative subset, 880 prompts):
238
+
239
+
240
+ | Category | SPEED-Bench Acceptance Rate |
241
+ | ------------------- | --------------------------- |
242
+ | coding | 3.46 |
243
+ | humanities | 2.80 |
244
+ | math | 3.29 |
245
+ | multilingual | 4.01 |
246
+ | qa | 2.94 |
247
+ | rag | 3.74 |
248
+ | reasoning | 3.31 |
249
+ | roleplay | 2.34 |
250
+ | stem | 2.88 |
251
+ | summarization | 3.54 |
252
+ | writing | 2.70 |
253
+ | **Overall Average** | **3.18** |
254
+
255
+
256
+
257
+
258
+ ## Ethical Considerations
259
+
260
+ NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. 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.
261
+
262
+ For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards below.
263
+
264
+ Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://app.intigriti.com/programs/nvidia/nvidiavdp/detail).
265
+
266
+ SUBCARDS:
267
+
268
+ ## Explainability
269
+
270
+
271
+ | Field | Response |
272
+ | -------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
273
+ | **Intended Task/Domain** | Few-step text-to-image generation. |
274
+ | **Model Type** | DMD2-distilled diffusion transformer (MMDiT) packaged as a full Diffusers pipeline. |
275
+ | **Intended Users** | Developers and researchers evaluating or integrating latency-sensitive image generation. |
276
+ | **Output** | Generated RGB image. |
277
+ | **Describe how the model works** | A Qwen2.5-VL text encoder converts the prompt to conditioning embeddings. A DMD2-distilled Qwen-Image transformer denoises image latents in four deterministic FlowMatch Euler steps, and the VAE decodes the final latents into an image. |
278
+ | **Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of** | Not Applicable |
279
+ | **Technical Limitations & Mitigation** | The model is designed for the packaged four-step schedule; changing the number of inference steps or scheduler trajectory may degrade image quality or produce unexpected results. It was distilled with English captions at 1024 × 1024, so performance with non-English prompts has not been evaluated and may be less reliable. The pipeline accepts other resolutions when the width and height are divisible by 16, but only 1024 × 1024 has been tested and quality may vary at other sizes. The model may inherit biases and visual artifacts from the base model and training data and does not include a safety checker; deployers should conduct use-case-specific testing and add appropriate safeguards. |
280
+ | **Verified to have met prescribed NVIDIA quality standards** | Yes |
281
+ | **Performance Metrics** | Image quality, prompt adherence, throughput, and end-to-end latency. |
282
+ | **Potential Known Risk** | This model can generate synthetic images and may produce content that is inaccurate, offensive, or otherwise inappropriate. Users should implement robust safety guardrails — including content filtering, abuse monitoring, and access controls — to reduce the risk of harmful outputs. Users are responsible for ensuring that their use of the model complies with all applicable laws and regulations, and for regularly reviewing and updating their guardrails as risks evolve. |
283
+ | **Licensing** | **Governing Terms:** Use of this model is governed by the NVIDIA Open Model Agreement([https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/])) .
284
+
285
+ **Additional Information:** [Apache License 2.0](https://choosealicense.com/licenses/apache-2.0/).) |
286
+
287
+
288
+
289
+
290
+ ## Bias
291
+
292
+
293
+ | Field | Response |
294
+ | ------------------------------------------------------------------------------------------- | -------- |
295
+ | **Participation considerations from adversely impacted groups in model design and testing** | None |
296
+ | **Measures taken to mitigate against unwanted bias** | None |
297
+ | **Bias Metric** | None |
298
+
299
+
300
+
301
+
302
+ ## Safety & Security
303
+
304
+
305
+ | Field | Response |
306
+ | -------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
307
+ | **Model Application Field(s)** | Text-to-image generation, creative-content prototyping, and research. |
308
+ | **Describe the life-critical impact (if present)** | Not Applicable |
309
+ | **Use Case Restrictions** | Use must comply with the **Governing Terms:** Use of this model is governed by the NVIDIA Open Model Agreement([https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/])) .
310
+
311
+ **Additional Information:** [Apache License 2.0](https://choosealicense.com/licenses/apache-2.0/) , and applicable law. Image editing, image understanding, high-stakes decisions, and illegal or abusive content generation are outside the intended scope. |
312
+ | **Model 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. |
313
+
314
+
315
+
316
+
317
+ ## Privacy
318
+
319
+
320
+ | Field | Response |
321
+ | ------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------- |
322
+ | **Generatable or reverse-engineerable personal data** | No |
323
+ | **Was consent obtained for any personal data used** | Not Applicable |
324
+ | **Personal data used to create this model** | None Known |
325
+ | **How often is the dataset reviewed** | Before Release |
326
+ | **Was data from user interactions with the AI model used to train the model** | No |
327
+ | **Is there provenance for all datasets used in training** | Yes |
328
+ | **Does data labeling comply with privacy laws** | Yes |
329
+ | **Is the data compliant with data-subject requests for correction or removal** | No, not possible with externally-sourced data. |
330
+ | **Applicable NVIDIA Privacy Policy** | [NVIDIA Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) |
331
+