GroundShot Keeps Characters Consistent Across Every Shot in an AI Video, No Retraining Required

Researchers at Zhejiang University, Fudan University, and Baidu published GroundShot in June 2026. The paper describes a framework that solves one of the most persistent practical problems in AI video production. A character or object that looks one way in shot one looks different in shot five. No model weights are changed. No fine tuning is required. GroundShot works on top of whatever video generation model you already use.

The Problem It Solves

When you generate a multi shot video in Wan 2.2, HunyuanVideo, or any other current model, each shot is generated largely in isolation. The model does not remember what your lead character looked like two shots ago. It generates a plausible looking person based on the text prompt. The result is a character whose face, clothing, and lighting shift from shot to shot in ways that read immediately as AI artifacts to any viewer.

This problem is called entity drift. It gets worse as the video gets longer. A two shot scene has one transition to manage. A ten shot sequence has nine. Each transition introduces new opportunities for the character to look differently. By the end of a scene, the person on screen may share almost nothing visually with the person who appeared at the beginning.

Entity drift is the main reason AI filmmakers cannot yet produce scenes at feature film length using current tools. Individual shots can look cinematic. Assembled sequences break the illusion.

What GroundShot Does

GroundShot builds a visual memory for each entity in the video. An entity is any recurring element. Named characters, specific objects, and locations that appear in multiple shots all count. As each shot is generated and accepted, GroundShot extracts visual information about every entity that appeared in it and stores that information in a per entity memory bank.

Before generating the next shot, the system retrieves the best available visual reference for every entity that will appear in that shot. It feeds those references into the generation process. The model generates the new shot with the existing appearances as a visual anchor, rather than working only from a text description.

The memory is not static. GroundShot verifies each generated shot before adding its entity information to the bank. Only shots that meet a reliability threshold contribute new data. This prevents a badly generated shot from corrupting the memory that all future shots depend on.

Shot Scheduling

The most technically distinctive part of GroundShot is how it decides which shot to generate first.

A conventional workflow generates shots in the same order they appear in the finished video. Shot one is generated first, then shot two, then shot three. That order feels natural but it creates a consistency problem. When you reach shot ten, you have only the reference images from shots one through nine to anchor the entities. If those early shots were generated before the memory was well established, the references may not be reliable.

GroundShot generates shots in the order that maximizes the quality of entity references, not the order the shots appear in the final video. It analyzes the full script before generation begins and calculates which shots will provide the most useful visual information for later shots. It generates those shots first, building a strong visual foundation before tackling shots where the entity reference demands are highest.

Once all shots are generated, GroundShot reassembles them into narrative order. The viewer sees them in story sequence. The generation happened in a different sequence optimized for consistency.

This approach, which the authors call quality aware shot scheduling, is what separates GroundShot from simpler approaches that pass entity images between shots without considering the order of generation.

The EntityBench Benchmark

The authors released EntityBench alongside the paper. It is a benchmark for measuring entity consistency in multi shot video generation at the level of individual characters and objects across shots, not just averages across a whole clip.

Prior evaluation approaches measured consistency across a whole clip but could not isolate whether a specific character remained consistent across the shots where they appeared. EntityBench measures this per entity and per shot. That precision matters for filmmakers because a six shot scene with three recurring characters requires consistent tracking of each character independently, not a single aggregate score for the sequence.

Results show measurable improvement in entity consistency across narrative distance, specifically the gap between a character's first appearance and their later appearances, compared to baseline generation without entity memory.

What Models It Runs On

GroundShot is designed to run on top of any existing video generation model without modifying the model's weights. The paper demonstrates results with current open source video models including Wan 2.2 and HunyuanVideo. Both are available through AI FILMS Studio.

The framework slots between your prompt and the generation model. You write the script. GroundShot parses it, schedules the shots, manages the entity memory, and calls the underlying model once per shot. The model generates each shot as normal. GroundShot handles the consistency logic around it.

This design means GroundShot does not become obsolete when better video models are released. Each new model it is paired with brings better base quality. The consistency layer on top remains the same.

Where This Fits in the Multi Shot Toolkit

MultiShotMaster addressed the problem of generating shots that match a specified duration and timing. ShotPlan provided a framework for structuring cinematic shot sequences from a script. GroundShot, published the same month, targets the third and hardest part of the problem. It keeps the same characters, the same objects, and the same world consistent across every cut.

The three tools address different constraints. ShotPlan handles shot structure. MultiShotMaster handles shot timing. GroundShot handles entity appearance. A filmmaker working on a short film in 2026 can use all three in sequence to produce something that was not practically achievable in 2025. A multi shot narrative with consistent characters, correct timing, and cinematic structure, generated entirely with AI video tools.

The same principle applies to any model that can generate from image and text inputs. Wan-Dancer-14B showed how Wan 2.2 can be extended for specialized creative tasks like music to dance generation. GroundShot extends it for narrative consistency across shots. Neither requires retraining the base model.

Code Status

The GroundShot paper was published in June 2026. No code repository has been released as of August 2026. The paper describes the method in sufficient technical detail for researchers to implement it, but tools accessible to filmmakers depend on a public code release or integration into existing generation pipelines.

The pattern for papers like this is a code release within one to three months of the paper, often accompanied by a project page and demonstration videos. When code ships, the most practical integration path for filmmakers will be through video generation tools that already support Wan 2.2 or HunyuanVideo, since GroundShot is designed to wrap those models rather than replace them.

The paper is at arXiv:2606.20799. The EntityBench companion benchmark is at arXiv:2605.15199.

---

Sources

arXiv "GroundShot: Visually Consistent Multi-Shot Long Video Generation via Entity-Grounded Shot Scheduling" (Yixuan Lai, Tianjia Shao, Kun Zhou, Weijia Dou, Siyu Zhu, Jingdong Wang. June 18, 2026)

arXiv (HTML) Full paper with figures (pipeline diagrams and qualitative results)

arXiv "EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation" (companion benchmark paper)

arXiv "ShotStream: Streaming Multi-Shot Video Generation for Interactive Storytelling" (related approach for context)

MultiShotMaster Project page (comparable multi shot system)

arXiv (PDF) Full paper PDF (benchmark tables and method details)