ubr-physical-ai/cosmos3-i2v-survival-sdg

Cosmos3 image-to-video: subject survival on synthetic rescue scenes

26 generated clips — 13 from Cosmos3-Super and 13 from Cosmos3-Nano, conditioned on the same 13 source frames — plus the per-frame scoring that produced them. Published as the raw material behind a negative result: image-conditioned generation does not hold a subject in place across a clip, and the bigger model does not fix it.

Provenance — read this first

Every clip is conditioned on a frame from NVIDIA Isaac Sim renders (sdg_clip_hall_buckets__orbit00..12), the same synthetic family as `ubr-physical-ai/isaac-sdg-rescue-target`.

No real footage was used, and no identifiable person appears in any source frame. A separate set of image-to-video clips derived from real footage exists and is deliberately not published, because consent from the people visible in it is unresolved. Nothing from that set is here.

What is in the box

i2v_super/    13 mp4   Cosmos3-Super
i2v_nano/     13 mp4   Cosmos3-Nano
survival/     2 csv    per-clip, per-frame judge output

All 26 clips share one source set, one prompt and one seed, so the two arms differ only by model.

The prompt names environment and lighting only — deliberately. It never names the camera platform, because naming it pulls the generator toward rendering one.

The measurement, with its N

A zero-shot detector judges whether the subject present in frame 0 is still present at frames 23, 46, 69 and 92. kept = 1 when the best box overlaps the frame-0 box above threshold.

Do not read these as a curve. Each cell is at most 13 clips, and the sequences are not monotone — Nano runs 1 → 4 → 1, Super 4 → 8 → 3. At this sample size a single clip moves a cell by 8 percentage points, and the differences between frame indices are not resolvable. Two clips are missing from the f46 row (n=11); that gap is unexplained and left visible rather than patched.

What the data does support, because it is the same at every index after f0: both models lose the subject almost immediately after the conditioning frame. Super retains it more often than Nano at every index, which is consistent with a larger model helping, but 13 clips cannot establish the size of that effect.

Why this exists

The intended use was domain enrichment for a search-and-rescue perception model: take a scene, generate variation, keep the subject. It does not work this way. An image-conditioned generator is anchored only at frame 0, so the subject drifts or disappears as soon as the generation moves away from it. Per-frame conditioning — a control signal re-applied at every frame — is the thing that holds a subject, not a bigger backbone.

These clips are published so that conclusion can be checked rather than taken on trust, and so the same experiment does not need re-running to reproduce the failure mode.

Limitations

• —13 source frames, one scene family. All sources are hall_buckets orbits. Nothing here generalises to other scenes, and the two arms share the same 13 conditioning frames.
• —Synthetic throughout. Isaac Sim renders, indoor, well lit. No real-world transfer claim is made.
• —The judge is zero-shot and was not itself validated on these clips; a missed detection and a lost subject are indistinguishable in this data.
• —Not an operational result. This is a bench experiment from a hackathon build, not evidence about any deployed system.

Licence

Clips and CSVs: CC-BY-4.0.

The clips are outputs of NVIDIA Cosmos3-Super and Cosmos3-Nano, which are distributed under OpenMDW 1.1. That licence states it "does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs generated by using the Model Materials", so the outputs here carry no obligation from it. Its attribution requirements attach to redistributing the model, which this dataset does not do.

Produced by Team UBR Stack for the NVIDIA / OpenHackathons / Oracle Open Models Codefest 2026.