Conversational Editing: Multi-Turn State Preservation
A specialized template for testing the model's ability to maintain narrative state across multiple sequential editing turns in a conversational interaction. This four-turn scaffold validates cumulative state preservation, instruction-native context understanding, and the absence of regression to prior states during iterative refinement of the same video content.
Stages
A young professional woman with dark brown hair (shoulder-length, slight wave), warm brown eyes, and a thoughtful expression. She wears a burgundy wool blazer with subtle texture, cream-colored blouse with small pearl buttons, and delicate gold jewelry. She is seated at a modern wooden desk in a bright office environment with soft afternoon sunlight streaming through tall windows. A potted plant is visible in the background. She gazes thoughtfully at a laptop screen, fingers poised over the keyboard.Establish the base character and scene identity. This becomes the anchor state for all subsequent edits.
Add a colleague (male, Asian, wearing charcoal dress shirt) entering from the left side of the frame. He carries a folder and appears engaged in conversation with the woman. Her facial expression changes to welcoming and professional. The desk environment remains unchanged. Lighting and color palette preserved.First edit: introduce a second character without disrupting the established scene state.
The colleague now sits in the visitor's chair to the right of the desk. Both characters are engaged in discussion — the woman leans forward slightly, gesturing with her right hand. The colleague nods. Their clothing details remain exactly as established. The office lighting and plant background are unchanged.Second edit: cumulative state — both characters must retain their identity, clothing, and the scene must accumulate changes rather than reset.
Remove the colleague. The woman returns to her original posture, gazing at the laptop. Her expression returns to thoughtful concentration. The desk, plant, lighting, window, and all scene details are exactly as they were in the base generation. No accumulated artifacts or changes from the colleague's presence remain.Critical test: state regression prevention. The model must remove the colleague without regressing to a cached intermediate state.
Compiled Prompt
role: senior_prompt_engineer @ omniveo
target_model: omni-2026.05
subject: <SUBJECT>
stage[1].generation := scene(character, outfit=burgundy_blazer, setting=office)
stage[2].turn_1 := add(colleague, position=left, interaction=conversation)
stage[3].turn_2 := modify(colleague, position=seated, interaction=discussion)
stage[4].turn_3 := remove(colleague, reset_to=base_state, no_artifacts=true)
assert: state_preservation_cumulative == true
assert: regression_prevention == true
assert: identity_consistency == true
assert: dialogue_coherence == true