Refine I2V 4-step model card overview

Summary: Add concise 4-step model overview with links, clarify prompt upsampling is optional for distilled models, and document the conservative PBR speedup basis. Validation: git diff --check.

README.md

@@ -22,6 +22,16 @@ tags:
 
 ## Description
 
+Cosmos3-Super-Image2Video-4Step is a 4-step distilled version of the base [Cosmos3-Super-Image2Video](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video) model. Given an input image plus an optional text prompt, it generates a temporally coherent video.
+
+- **What's different from the base model:** Distilled into a fixed 4-step sampling schedule without the need for classifier-free guidance. At the recommended 50-step + CFG setting for the Image2Video model, this replaces 50 × 2 = 100 model evaluations with 4, leading up to 25× fewer diffusion-model evaluations per video (see [PBR](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step#pbr-performance-benchmark-reporting) for latency estimates).
+- **Quality:** Ranked #1 among open-source models on the Artificial Analysis Image-to-Video leaderboard (as of 2026/07/17; see [Benchmarks section](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step#benchmarks)).
+- **Input / Output:** Image(s) in → MP4 video out. Recommended 480p (832×480, 16:9); 189 frames by default.
+- **Example usage and output:** See [Usage](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step#usage-run-inference-with-vllm-omni) section.
+- **Hardware:** The 64B model needs a multi-GPU H100/H200 node (4–8 GPUs) or a single B200; it does not fit on a single smaller GPU. See [Usage](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step#usage-run-inference-with-vllm-omni). Full model details below.
+
+Cosmos3-Super-Image2Video-4Step was developed by NVIDIA as a part of Cosmos3.
+
 Cosmos3 is a collection of Omnimodal world models capable of generating dynamic, high-quality video, image, audio, and action commands from combinations of text, image, video, and action trajectory inputs. It serves as a foundational building block for a broad range of Physical AI applications and research spanning world understanding, world generation, simulation, and embodied policy learning.
 
 This model is ready for commercial and non-commercial use.
@@ -333,12 +343,13 @@ Latency is reported in seconds. The suffix in each resolution column indicates t
 5. vLLM-Omni source results report "Avg. Generation Time" for the corresponding vLLM-Omni inference configuration.
 6. Cells marked `N/A` indicate configurations unavailable because the 64B model does not fit on a single GPU. Cells marked `Not tested` indicate configurations for which no source measurement was reported.
 7. All displayed values were calculated as `Cosmos3-Super source latency ÷ 17.5` and rounded to three decimal places.
+8. Note on speedup basis: the 17.5× above is extrapolated from a 35-step base, which was the step count used in the original Cosmos3-Super base model speed benchmark. For Image-to-Video specifically, the recommended sampler setting is 50 steps, corresponding to `50 × 2 ÷ 4 = 25×` fewer model evaluations; this is the figure used in general reporting. The 17.5× estimate is retained here as the more conservative, benchmark-matched number.
 
 </sub>
 
 ## Usage: Prepare Prompts with Prompt Upsampling
 
-`Cosmos3-Super-Image2Video-4Step` supports dense natural-language prompts together with a first-frame image. It also supports JSON-format upsampled prompts, which can further improve generation quality. Description and code can be found [here](https://github.com/nvidia/cosmos-framework/blob/main/docs/prompt_upsampling.md).
+`Cosmos3-Super-Image2Video-4Step` supports dense natural-language prompts together with a first-frame image. Prompt upsampling is optional for distilled models: you can run inference directly with a natural-language prompt, or use a JSON-format upsampled prompt for potential quality improvements. Description and code can be found [here](https://github.com/nvidia/cosmos-framework/blob/main/docs/prompt_upsampling.md).
 
 To create an upsampled prompt using Opus-4.7: