Finalize Qwen-Image-Flash model card

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/qwen-image-flash-showcase.jpg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -21,9 +21,13 @@ tags:
 
 ## Description:
 
-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.
+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 distillation used DMD2 from [NVIDIA FastGen](https://github.com/NVlabs/FastGen), [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.
 
-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.
+## Showcase
+
+![Qwen-Image-Flash showcase](./assets/qwen-image-flash-showcase.jpg)
+
+All showcase images were generated with [nvidia/Qwen-Image-Flash](https://huggingface.co/nvidia/Qwen-Image-Flash).
 
 This model is ready for commercial or non-commercial use.
 
@@ -94,9 +98,10 @@ Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated sys
 
 **Supported Runtime Engine(s):**
 
-- [Hugging Face Diffusers](https://github.com/huggingface/diffusers) 0.38.0
-- [PyTorch](https://pytorch.org/)
-- [Transformers](https://github.com/huggingface/transformers) 5.5.0
+- [Hugging Face Diffusers](https://github.com/huggingface/diffusers) 0.38.0 with Transformers 5.12.1
+- [SGLang Diffusion](https://github.com/sgl-project/sglang): `lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4`
+- [vLLM-Omni](https://github.com/vllm-project/vllm-omni): `vllm/vllm-omni:v0.24.0`
+- [TensorRT-LLM VisualGen](https://github.com/NVIDIA/TensorRT-LLM): `nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc21`
 
 **Supported Hardware Microarchitecture Compatibility:**
 
@@ -139,16 +144,16 @@ Qwen-Image-Flash was distilled with NVIDIA Model Optimizer 0.45.0.
 
 ## Evaluation Dataset:
 
-**Link:** MTBench, for more details, see [here](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge)
-**Data Collection Method by dataset:** Hybrid: Manually-collected, Synthetic
-**Labeling Method by dataset:** Hybrid: manually-labelled, Synthetic
-**Properties:** 3,300 multi-turn dialogue sequences.
+**Link:** [Qwen-Image-Bench](https://arxiv.org/abs/2605.28091), [OneIG-Bench-EN](https://oneig-bench.github.io/)
+**Data Collection Method by dataset:** Manually-collected
+**Labeling Method by dataset:** Qwen-Image-Bench: Automated (Q-Judger with professional human supervision); OneIG-Bench-EN: Automated benchmark metrics
+**Properties:** Qwen-Image-Bench: 1,000 prompts; OneIG-Bench-EN: 1,120 English prompts across five categories.
 
 ## Inference:
 
-**Acceleration Engine:** Hugging Face Diffusers with PyTorch
+**Acceleration Engine:** Hugging Face Diffusers, SGLang Diffusion, vLLM-Omni, and TensorRT-LLM VisualGen
 
-**Test Hardware:** NVIDIA Hopper H100
+**Test Hardware:** NVIDIA GB200
 
 ## DMD2 Distillation
 
@@ -158,102 +163,101 @@ The teacher target used classifier-free guidance (CFG) 4.0 during distillation.
 
 ## Usage
 
-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:
+### Diffusers
 
-```bash
-pip install "diffusers==0.38.0" "transformers==5.5.0" accelerate safetensors
-```
+**Tested container:** `vllm/vllm-omni:v0.24.0` (`diffusers==0.38.0`, `transformers==5.12.1`)
 
 ```python
 import torch
 from diffusers import QwenImagePipeline
 
-model_id = "nvidia/Qwen-Image-Flash"
-
 pipe = QwenImagePipeline.from_pretrained(
-    model_id,
+    "nvidia/Qwen-Image-Flash",
     torch_dtype=torch.bfloat16,
 ).to("cuda")
 
-prompt = (
-    'A cinematic photograph of a small coffee shop. A chalkboard clearly reads '
-    '"Qwen Coffee — open in a flash", detailed natural lighting, Ultra HD, 4K.'
-)
-
 image = pipe(
-    prompt=prompt,
+    prompt="A red fox in a snowy pine forest at golden hour, photorealistic, sharp focus, soft bokeh",
     width=1024,
     height=1024,
     num_inference_steps=4,
     true_cfg_scale=1.0,
+    guidance_scale=None,
+    negative_prompt=None,
     generator=torch.Generator(device="cuda").manual_seed(42),
 ).images[0]
 
-image.save("qwen-image-flash.png")
-```
-
-For lower accelerator-memory use, construct the pipeline without `.to("cuda")`, then enable model CPU offload:
-
-```python
-pipe = QwenImagePipeline.from_pretrained(
-    model_id,
-    torch_dtype=torch.bfloat16,
-)
-pipe.enable_model_cpu_offload()
+image.save("qwen-image-flash-diffusers.png")
 ```
 
-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.
-
-**Required Inference Settings:**
-
-- Use `num_inference_steps=4`.
-- Use `true_cfg_scale=1.0` and omit `negative_prompt`; the student has already internalized the teacher's CFG target.
-- Keep the packaged static shift-3 scheduler and do not pass custom `sigmas`.
+### SGLang Diffusion
 
+**Tested container:** `lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4`
 
+```bash
+docker run --rm --gpus all --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 -e HF_TOKEN -v qwen-image-hf-cache:/root/.cache/huggingface -v "$PWD:/workspace" -w /workspace lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 sglang generate --model-path nvidia/Qwen-Image-Flash --prompt "A red fox in a snowy pine forest at golden hour, photorealistic, sharp focus, soft bokeh" --width 1024 --height 1024 --num-inference-steps 4 --guidance-scale 1.0 --true-cfg-scale 1.0 --seed 42 --save-output --output-file-path qwen-image-flash-sglang.png
+```
 
-## Evaluation
+### vLLM-Omni
 
-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.
+**Tested container:** `vllm/vllm-omni:v0.24.0`
 
-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).
+```bash
+# Server
+docker run --rm --gpus all --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 -e HF_TOKEN -p 8091:8091 -v qwen-image-hf-cache:/root/.cache/huggingface vllm/vllm-omni:v0.24.0 vllm serve nvidia/Qwen-Image-Flash --omni --host 0.0.0.0 --port 8091
+```
 
-Acceptance length on [MT-Bench](https://huggingface.co/spaces/lmsys/mt-bench) (80 prompts):
+```bash
+# Client
+curl -fsS http://127.0.0.1:8091/v1/images/generations -H 'Content-Type: application/json' -d '{"prompt":"A red fox in a snowy pine forest at golden hour, photorealistic, sharp focus, soft bokeh","n":1,"size":"1024x1024","response_format":"b64_json","num_inference_steps":4,"true_cfg_scale":1.0,"guidance_scale":1.0,"seed":42}' -o qwen-image-flash-vllm-omni-response.json
+set -o pipefail; jq -er '.data[0].b64_json' qwen-image-flash-vllm-omni-response.json | base64 --decode > qwen-image-flash-vllm-omni.png
+```
 
+### TensorRT-LLM
 
-| Category            | MT Bench Acceptance Rate |
-| ------------------- | ------------------------ |
-| writing             | 2.93                     |
-| roleplay            | 2.54                     |
-| reasoning           | 3.09                     |
-| math                | 4.54                     |
-| coding              | 3.42                     |
-| extraction          | 4.13                     |
-| stem                | 2.77                     |
-| humanities          | 2.43                     |
-| **Overall Average** | **3.23**                 |
+**Tested container:** `nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc21`
 
+```python
+from tensorrt_llm import VisualGen
+
+
+def main():
+    visual_gen = VisualGen(model="nvidia/Qwen-Image-Flash")
+    try:
+        params = visual_gen.default_params
+        params.width = 1024
+        params.height = 1024
+        params.num_inference_steps = 4
+        params.guidance_scale = 1.0
+        params.negative_prompt = None
+        params.seed = 42
+
+        output = visual_gen.generate(
+            inputs="A red fox in a snowy pine forest at golden hour, photorealistic, sharp focus, soft bokeh",
+            params=params,
+        )
+        output.save("qwen-image-flash-tensorrt-llm.png")
+    finally:
+        visual_gen.shutdown()
+
+
+if __name__ == "__main__":
+    main()
+```
 
-Acceptance length on [SPEED-Bench](https://huggingface.co/datasets/nvidia/SPEED-Bench) (qualitative subset, 880 prompts):
 
 
-| Category            | SPEED-Bench Acceptance Rate |
-| ------------------- | --------------------------- |
-| coding              | 3.46                        |
-| humanities          | 2.80                        |
-| math                | 3.29                        |
-| multilingual        | 4.01                        |
-| qa                  | 2.94                        |
-| rag                 | 3.74                        |
-| reasoning           | 3.31                        |
-| roleplay            | 2.34                        |
-| stem                | 2.88                        |
-| summarization       | 3.54                        |
-| writing             | 2.70                        |
-| **Overall Average** | **3.18**                    |
+## Evaluation
 
+The following results compare Qwen-Image-Flash with Qwen-Image and Qwen-Image-Lightning:
 
+Evaluation was performed at 1024 × 1024 resolution on an NVIDIA B200 GPU.
 
+| Model | [Qwen-Image-Bench](https://arxiv.org/abs/2605.28091) | [OneIG-Bench-EN](https://oneig-bench.github.io/) |
+| --- | ---: | ---: |
+| Qwen-Image | 49.070 | 0.886 |
+| [Qwen-Image-Lightning](https://huggingface.co/lightx2v/Qwen-Image-Lightning) | 46.980 | 0.887 |
+| **Qwen-Image-Flash (Ours)** | **47.080** | **0.885** |
 
 ## Ethical Considerations