Add image-to-video with negative prompt example

Adds an 'Image to video with negative prompt' usage example that mirrors the existing image-to-video section, reusing the same input image and prompt but passing the curated assets/negative_prompt.json as the per-sample negative prompt (JSON-encoded string, matching the model's default negative caption format). Includes the generated example output (assets/edge_i2v_negprompt_output.mp4, 832x480, 121 frames @ 24fps).

.gitattributes

@@ -51,3 +51,4 @@ assets/edge_i2v_output.mp4 filter=lfs diff=lfs merge=lfs -text
 assets/edge_action_fd_umi_2chunk_output.mp4 filter=lfs diff=lfs merge=lfs -text
 assets/edge_action_id_av_0_output.png filter=lfs diff=lfs merge=lfs -text
 assets/edge_action_id_av_1_output.png filter=lfs diff=lfs merge=lfs -text
+assets/edge_i2v_negprompt_output.mp4 filter=lfs diff=lfs merge=lfs -text

README.md

@@ -514,6 +514,49 @@ Example output:
 
 <video controls width="832" height="480" src="https://huggingface.co/nvidia/Cosmos3-Edge/resolve/main/assets/edge_i2v_output.mp4"></video>
 
+#### Image to video with negative prompt
+
+The same image-to-video setup as above, but with a structured **negative prompt** to steer the model away from common failure modes (temporal flicker, morphing geometry, broken physics, unstable camera push-in). We reuse the same input image and prompt, and pass the curated `negative_prompt.json` from `assets/` as the per-sample `negative_prompt` (JSON-encoded as a string, mirroring the model's default negative caption format).
+
+```shell
+BASE="https://huggingface.co/nvidia/Cosmos3-Edge/resolve/main/assets"
+python - "$BASE" <<'PY'
+import json, sys, urllib.request
+base = sys.argv[1]
+prompt = urllib.request.urlopen(f"{base}/example_i2v_prompt.json").read().decode()
+negative = urllib.request.urlopen(f"{base}/negative_prompt.json").read().decode()
+json.dump(
+    {
+        "model_mode": "image2video",
+        "prompt": prompt,
+        "vision_path": f"{base}/example_i2v_input.jpg",
+        "negative_prompt": json.dumps(json.loads(negative)),
+    },
+    open("i2v_edge_neg.json", "w"), indent=2,
+)
+PY
+
+python -m cosmos_framework.scripts.inference \
+    --parallelism-preset=latency \
+    -i i2v_edge_neg.json \
+    -o outputs/omni_edge_neg \
+    --checkpoint-path Cosmos3-Edge \
+    --resolution 480 \
+    --num-frames 121 \
+    --fps 24 \
+    --num-steps 50 \
+    --guidance 5.0 \
+    --shift 3.0 \
+    --sampler unipc \
+    --seed 0
+```
+
+The generated video is written to `outputs/omni_edge_neg/<sample_name>/vision.mp4`.
+
+Example output:
+
+<video controls width="832" height="480" src="https://huggingface.co/nvidia/Cosmos3-Edge/resolve/main/assets/edge_i2v_negprompt_output.mp4"></video>
+
 #### Action — forward dynamics
 
 Forward dynamics rolls out one action chunk at a time. This self-contained script fetches the UMI conditioning frame and action chunks, runs `cosmos_framework.scripts.inference` per chunk, feeds each chunk's final generated frame in as the next chunk's conditioning image, and stitches the 2-chunk rollout. Install `imageio[ffmpeg]` first.

assets/edge_i2v_negprompt_output.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2d41ae997adea1eef0f046b837404d6add2798e050974ea36696cee4b3efd00
+size 23017533