AI Video Still Relies on “Gacha”: Usability and Rework Are the Real Bottlenecks

Introduction

Generating an AI video may take only a few minutes. Producing a shot that actually matches a creator’s requirements is another matter. The repeated cycle of generation and selection is often described as “gacha.” According to the source interviews, an AI short animation may require roughly ten generation attempts from storyboard to final cut before one satisfactory result emerges. Faster generation, therefore, does not automatically mean faster delivery.

The real problem: fast generation is not usable content

Commercial production demands more than visually convincing frames. E-commerce images must preserve a product’s shape, color, and brand marks while producing multiple formats. Video projects must maintain character identity, movement, spatial relationships, and shot transitions. A single inconsistent shot can trigger additional revisions across the sequence.

Long-form narrative video is particularly difficult. The source cites research arguing that many current workflows lack a structured layer between scripts and video models. A script may describe what happens, but not specify camera movement, character placement, or which details must remain stable across shots. If these constraints are not represented before generation, creators are forced to repair problems afterward.

The cost is more than model inference

If one generation costs 10 yuan and only one in ten attempts is usable, generation alone costs 100 yuan per usable shot. At a 50% usable rate, the average cost falls to about 20 yuan. Human review, editing, waiting time, and post-production add further expense. Improving usable output can therefore matter as much as reducing the price of an individual API call.

The material describes three ways that Tencent Cloud’s WAND is trying to reduce rework:

• Adding constraints before generation: 3D block models can define layout, object positions, and some motion relationships before a video model renders the final scene.
• Tuning for specific industries: E-commerce prioritizes product fidelity, while short-form drama emphasizes character continuity, motion, and shot transitions.
• Repairing local defects after generation: If the main image is acceptable and only subtitles or small regions are wrong, localized erasure and reconstruction may be cheaper than regenerating the entire clip.

Company-reported cases cited in the material say that Man Sen Culture raised shot usability from about 30% to 50%–80% and reduced the cost of premium AI short dramas by around 30%. ZiXun Technology reported that the effective use rate of e-commerce image assets rose from 50%–60% to 70%–80%. These are vendor or customer cases, and results will vary by workflow and task.

What agents can—and cannot—do

A complete production pipeline may involve generation, repair, enhancement, voice, subtitles, and editing. Fixed workflows work well for predictable tasks, while agents can select tools, arrange steps, and initiate model calls from natural-language requests. One platform cited in the material reported that agents initiated 82% of its video-model calls.

Automation, however, is not the same as autonomous delivery. A professional creator may want to change only a facial expression or camera movement, not regenerate the entire frame. Without intermediate results and localized editing, greater automation can reduce control rather than improve it.

Why it matters

The competitive frontier in AI video is moving from isolated model quality to the complete production system. Creative intent must be converted into explicit constraints, multiple tools must be coordinated, and the final result must remain repairable and editable. The most valuable products will not simply generate images faster; they will deliver usable footage with fewer retries while keeping creators in control.

Source: InfoQ Chinese