Qualitative editing examples
Action substitution
The requested change replaces walking toward the door with drinking water. In the shown Gemini Omni example, drinking is added near the end while much of the walking remains. Check both the presence of the target event and removal of the source event. View the example.
Clothing edits across turns
The sequence edits clothing colors, preserves the edits in the next turn, then restores both colors. It does not test undoing only one change while preserving another. View the sequence.
Dog–bone interaction
The source contains a yellow-brown bone on the ground. The archived target has a differently shaped white bone in the mouth from the first frame, so it does not demonstrate the original bone being picked up. The third clip is LucyEdit: the original bone stays on the ground and the dog's coat changes. The fourth clip, newly supplied by the presenter, is Gemini Omni Flash: the dog approaches the original bone near the end, but the clip does not show it lifting and carrying the bone. Attribution follows the presenter's correction. These four clips belong to one example. Use Play all to start the comparison together and Pause all to pause it. View the comparison.
These examples diagnose event timing, object continuity, interaction execution, and preservation. They do not establish overall model rankings or failure rates.
Evaluation snapshot — 2026-09-26
The three results shown in the deck use different metrics and protocols. Each is a diagnostic of the judge, and their scales must not be combined.
| Subset | Judge and protocol | No-edit score | Interpretation |
|---|---|---|---|
| VEFX Instance Motion33 | VEFX-Reward-4B, official duration-dependent sampling at 4 frames/s; IF, 1–4 | 2.7089 | Above the tested editing outputs under this protocol. |
| Omni Temporal-200 | Qwen3-VL-32B with the Omni rubric and all decoded video frames; Overall, 0–100 | 33.340 | Above editing outputs with results available under the full-video protocol. |
| IVEBench Subject Motion Short64 | Qwen2.5-VL-72B with official IS preprocessing; IS, 1–5 | 1.6875 | This setup gives the unedited output a low instruction score. |
Source-as-output deliberately performs no edit and is excluded from the editor ranking. High preservation or smoothness cannot establish action completion. In the Omni sensitivity study, switching the same Qwen3-VL-32B judge from 12 sampled frames to all decoded frames changes the no-edit Overall score from 17.608 to 33.340. The direction of change also depends on judge size, so more frames are not automatically a more accurate evaluation.
Local configuration names
FM-only denotes a Bernini-R 1.3B adapter trained with direct flow-matching supervision and source-plus-instruction inference. Zebra denotes a reference-conditioned controller whose pipeline additionally generates a pseudo-reference. Neither name denotes an independently published paper. Their extra inputs and generation cost must be reported when comparing results; neither has a full-video record in the frozen Omni comparison used here.
These values are transcribed from the presenter's frozen 2026-09-26 aggregate report used to prepare the deck. The public page documents the displayed results and protocol limits; it is not a raw benchmark release or a new evaluation run.