Development summary
Indriya’s “Making AI Work” video programme provides an experimental environment for studying controllability and reliability in generated audiovisual sequences. Iterative development has investigated character consistency, object continuity, action timing and selective correction of generated content.
The work has produced prototype sequences, documented failure cases, local repair implementations and verification procedures. AI-assisted workflows support asset preparation, editing, assembly and quality checks. Humans review whether a sequence meets the intended creative requirements.
These assets support further systematic evaluation. Comparative performance, generalization across scenes and automated quality assessment remain research objectives.
1. Turn the scene into testable requirements
A narrative scene contains constraints beyond its style. The same character must remain recognizable. An object count must remain stable. Contact between a paw and a pan must make physical sense. An action may need to finish before a character reacts.
The existing sequences have exposed failures involving appearance, object changes, contact and event ordering. Those cases give the proposed evaluation concrete requirements rather than a general instruction to produce a better video.
The next comparison would test ordinary prompting against structured shot requirements, approved reference assets and explicit start and end states. It would hold the model, shot briefs and generation budget constant where possible, then assess constraint violations, accepted-shot rate and attempts needed for acceptance. The matched comparison is the next step in the research.
2. Repair selected defects
A full regeneration can alter content that a reviewer has already accepted. Selective correction addresses a narrower region or defect while seeking to retain the rest of the shot, its timing and its existing audio. Local repair implementations form part of the current prototype work.
The research question is whether that narrower intervention can preserve approved content more reliably and with less effort than regeneration. A broader evaluation would compare local repair, reference-guided regeneration and whole-shot regeneration across matched defects.
A correction should be assessed for its intended effect and for unintended changes outside the target region. Temporal artifacts, human review time and cost per accepted correction also matter. A repair that looks correct in one frame can still fail across the sequence.
3. Align speech, expression and action
The programme includes dialogue preparation, timing against picture cuts and planning for lip-sync pilots. Speech, expression and action create related constraints: the line must occur at the intended point, the character should retain its appearance, and the reaction should belong to the scene.
Physical and temporal continuity are evaluated together. The required order of actions can be wrong even when each individual frame looks plausible. Human approval of a timing change does not settle remaining visual or lip-sync work.
Further investigation would use repeatable scene requirements and review criteria for alignment. Learning through short narrative video is a possible study direction, with comprehension and retention requiring their own viewer evaluation.
4. Human-supervised production agents
Agents and AI-assisted engineering support asset preparation, reference handling, version tracking, clip assembly and quality checks. The research question is how those workflows respond to feedback, detect a failure, select a suitable correction and preserve approved content.
A proposed comparison would match manual and agent-assisted tasks. It would record human interventions, failed attempts, total cost, elapsed time and final quality. Including unsuccessful runs is necessary to assess the process rather than only the accepted output.
Creative acceptance remains a human decision. Technical checks such as frame decoding and audio preservation support review but do not establish that a character, movement or expression is correct. The workflow needs both kinds of evidence.
5. Relate the task to established evaluation
VBench provides evaluation dimensions for generated video, including subject consistency, temporal flicker and motion smoothness. T2V-CompBench examines compositional problems such as object interactions, spatial relationships, action binding and object counts. These are relevant background for the failures encountered in narrative sequences.
An Indriya-specific evaluation would also need scene-level acceptance criteria and checks around local repair. General video scores cannot by themselves establish whether the required action order was preserved or an approved audio track remained unchanged.
The proposed measures therefore combine established video-quality dimensions with the constraints of each task. Benchmark scoring will need to sit alongside task-specific acceptance checks in a future comparison.
6. From case studies to broader comparisons
The current prototypes make the failure cases concrete and supply methods to investigate. A systematic comparison is the next step toward establishing whether those methods transfer beyond the individual reel.
We will compare repair methods across matched defects, recording correction effort, unintended changes and final sequence quality.
The programme’s current stage is exploratory prototype development. Comparative testing will examine structured references, scene constraints and targeted repair, with limits and unsuccessful cases reported alongside accepted results.
Multimodal models for sequence assessment
Nemotron 3 Nano Omni and Google Gemma models provide candidates for inspecting sampled frames and, where supported, accompanying audio. Our ongoing Apple AFM 3 work explores a separate platform-based route. The research question is whether these systems identify scene-constraint violations that human reviewers also find.
Compare model judgments against annotated defects and accepted sequences. Report missed defects, false alarms and temporal sampling. Apple ADM 3 Cloud supports image generation and editing.
References and publication status
References checked on 14 October 2026.