AI Harness in AI Animation

1. Introduction

1.1 Background

AI animation is the direction within the creative industry group with the highest technical threshold and the lowest degree of standardization. It must first be stated honestly: as of this document's retrieval baseline (2026-09-12), no dedicated national standard or industry-standard number for AI animation was found. Currently, only two types of authoritative material can be cited: local science-and-technology program results (Beijing Municipal Science and Technology Commission project Z241100007624006) and corporate public releases at international academic conferences (Tencent Games VISVISE, SIGGRAPH Asia 2025).

The direct consequence of this "standard vacuum" is: quality assessment lacks a unified quantitative metric. In the intelligent manufacturing direction, yield rate and missed-detection rate have industry-consensus algorithms; in the AI animation direction, how to compute character consistency, how to measure foot-sliding jitter, and how to judge 4K upscaling quality currently have only proposed recommended values but no standard thresholds. This is a reality that Harness construction in this direction must first face squarely.

The core technical difficulties have been clearly identified. The three major technical difficulties summarized by the Beijing Municipal Science and Technology Commission project are representative:

1. Character consistency control: character drift in long-sequence generation.
2. High-quality image generation: motion fluency and detail texture are hard to guarantee simultaneously.
3. Cross-scene coherence: coherent motion across consecutive frames and scenes.

All three point to the same thing — state maintenance across frames, across shots, and across scenes.

1.2 Definition and Scope

The AI Harness of this direction refers to: in the AI animation production scenario, an engineering hosting layer that takes on stages such as skeleton generation, skinning, animation generation, in-betweening, upscaling, and color grading, organizing scattered generative models into a controllable, reproducible, regression-verifiable animation production pipeline.

1.3 Positioning in the AI Harness System

图 1-1｜AI 动画在 AI Harness 六层体系中的定位（L1–L6）

数据来源：基于本文分析绘制的示意图。

Leading layers: L3 Orchestration & Control + L4 Memory & State. Bottleneck layer: L4 Memory & State.

Compared with the AI web-series direction, the L4 bottleneck of AI animation differs in two points:

1. Drift occurs at a finer physical scale: web-series drift is at the "face-swap level," while animation drift is at the "frame level" — foot sliding, jitter, and distortion are all millisecond-scale phenomena, making manual frame-by-frame inspection extremely costly.
2. State coupling introduced by two-stage rendering: the Beijing Municipal Science and Technology Commission project uses two-stage rendering (first motion and movement trajectories, then skin texture, clothing texture, lighting and shadow). Feature propagation must be maintained between the two stages, otherwise the second stage breaks the motion consistency of the first.

Tencent VISVISE's data confirms both points: compared with Alpha, MotionGen Beta reduced the probability of anomalies such as foot sliding, jitter, and distortion by 6 times, and MIB achieves 200 frames generated in 4 seconds. These are not "image quality improvements" but concrete measurements of consistency engineering.

1.4 Industry Status and Publicly Disclosed Quantitative Metrics

2. Glossary

3. Case Studies

3.1 Tencent Games VISVISE: Industry's First Full-Pipeline AI 3D Character Animation Pipeline

3.1.1 Background

Game industrialization has an enormous demand for 3D character animation with extremely high quality requirements: bone binding and skinning of a single 3A character traditionally require days of manual work, and motion creation requires repeated adjustment to eliminate foot sliding and jitter. In the traditional pipeline, the four stages of skeleton, skinning, animation, and in-betweening are completed by different tools and different people, and state is transferred between stages through file exchange — a primary source of quality loss.

3.1.2 Approach

At SIGGRAPH Asia 2025 (Hong Kong, December 15–18), Tencent Games released VISVISE, positioned as the industry's first AI full-pipeline 3D character animation production pipeline, comprising four major modules:

Decomposition from a Harness perspective:

• L2: The four modules are four registered tools, invoked in sequence by stage; each module has its own parameter scale and input specification.
• L3: Skeleton → skinning → animation → in-betweening is a strictly layered pipeline; the output of one stage is a hard constraint for the next.
• L4: MIB's autoregressive framework and "large-scale pretrained motion and transition representations" are essentially persisting motion state as a reusable representation, rather than remembering it temporarily within a session.
• L5: Using "probability of anomaly occurrence" as the core metric (reduced by 6 times), rather than subjective metrics such as "visual-quality score."

3.1.3 Results

• Production efficiency improved by more than 8x.
• Applied to nearly 100 game projects, including Peace Elite, Honor of Kings, PUBG Mobile, Teamfight Tactics, and League of Legends: Wild Rift.
• Skeleton generation compressed from days to minute-level; skinning automation rate 90%+.
• Academia and IP: multiple results published at GDC, SIGGRAPH, ICLR, devcom, among others; holds dozens of patents.

Source: Tencent Games, SIGGRAPH Asia 2025 (talk: VISVISE AI animation lead Zeng Zijiao, "Giving Geometry Life: A Full-Pipeline AI Pipeline for 3D Character Animation"); ifeng.com report, 2025.

3.2 The Beijing Municipal Science and Technology Commission AI Full-Pipeline Animation Production Platform Project

3.2.1 Background

Domestically, the production cycle of a typical single animation work is 2~3 years, much of which is consumed by consistency correction: the character looks different across shots, cross-scene motion is incoherent, and detail and motion are hard to balance. This is a structural bottleneck constraining the output capacity of domestic animation.

3.2.2 Approach

Science and technology plan project "Research, Development and Application of an AI Full-Pipeline Animation Production Platform" (project number Z241100007624006) of the Beijing Municipal Science and Technology Commission and the Zhongguancun Science Park Administrative Committee, implemented by Beijing Guanzhong Huyu Culture Technology Co., Ltd., with a period of 2024-09 to 2025-12. The three major technical difficulties and their solutions:

Difficulty 1 · Character consistency control

• Split the AI generation pipeline and strengthen input–output control to build an animation coherence control network.
• Add human-in-the-loop nodes within the generation pipeline.
• Precise instruction design realizes the transformation of "vague commands into precise content."
• Integrate temporal memory module, cross-frame attention, and feature propagation mechanisms to precisely control character appearance and motion trajectories across frames, solving character drift in long-sequence generation.

Difficulty 2 · High-quality image generation

• Two-stage rendering: the first stage renders motion and movement trajectories (preserving fluency); the second stage depicts skin texture, clothing texture, and lighting/shadow (preserving detail and ensuring cross-frame stability).
• At the video output stage, an super-resolution AI model intelligently raises the generated size to 4K level.

Difficulty 3 · Cross-scene coherence

• Integrate DiT global spatiotemporal modeling and 3D VAE to build an end-to-end solution of global modeling plus detail completion.
• The global attention mechanism models consecutive video frames as a whole.
• 3D VAE achieves feature-dimension compression and interpolation to alleviate frame jitter.
• The motion-feature binding technology of Motion LoRA strengthens cross-scene motion coherence.

From a Harness perspective, the most notable design of this project is the "addition of human-in-the-loop nodes in the generation pipeline" — which directly corresponds to the requirement in the group-level AGENTS.md that "consistency-critical nodes must be confirmed by a human up front." It is not a flaw but a correct engineering choice under current technological conditions.

3.2.3 Results

• Compressed the production cycle of a single animation work, typically 2~3 years in the domestic and international industry, to 5~6 months.
• Video output supports upscaling to 4K level.

Source: Beijing Municipal Science and Technology Commission results database (project number Z241100007624006), 2024–2025.

3.3 Engineering Practice and Reference Metrics for Character Consistency

Important note: the source of this case study is aggregated content without a URL; all values herein are proposed recommended values or industry estimates, not standard thresholds, and must not be cited as authoritative metrics. The value of retaining this case study is that it provides an actionable engineering framework and a toolchain checklist.

3.3.1 Background

In the absence of industry standards, practitioners must build their own consistency assessment system. This raises two questions: where the thresholds come from, and how the toolchain is combined.

3.3.2 Approach

The three-layer architecture of "feature anchoring — multimodal constraint — temporal coherence":

Toolchain (aggregated source): Stable Diffusion WebUI + LoRA / ControlNet, Leonardo AI, Midjourney (gen_id lock), Animate Anyone, ContextAnyone, Amazon Nova Canvas, ADetailer, RIFE, DaVinci Resolve.

There is also an Amazon Cloud FuzzyPixel team case: by fine-tuning the Nova Canvas model to tailor a consistency model for the protagonist of the animated work Picchu, training was completed in 12 hours (source is URL-less aggregated content).

3.3.3 Results (proposed recommended values, not standard thresholds)

Cost range (industry estimate, not official data): model and tool licensing 50–200 万元; local GPU cluster (8+ A100 GPUs) hardware + operations 120–200 万元/year; cloud compute for theatrical feature films 80–150 万元; workforce of 10–20 people for 3–6 months 100–300 万元; single theatrical-grade character-consistency-related cost approximately 380–800 万元; compared with the traditional approach, 30%–50% of visual-effects and post-production costs can be saved; storyboard production time is shortened from 2 weeks to 4 hours.

Source: aggregated content (no URL). All of the above values must be labeled as "proposed recommended values" or "industry estimates."

4. Practice Standards

4.1 AGENTS.md Specification

4.1.1. AGENTS.md (AI Animation direction)

# AGENTS.md —— AI 动画（AI Animation）

## 角色与边界
- 你运行在 AI 动画制作管线之上，负责骨骼生成、蒙皮、动画生成、插帧、超分与调色六类任务。
- 你负责执行与制作，不负责艺术方向与表演判断。艺术方向由动画导演与美术总监决定。
- 你不得在无具名人类动画导演签核的情况下输出成片序列。
- 你不得生成受版权保护的既有动画形象，除非持有书面授权。
- 你不得使用未授权的真人肖像作为角色参考。

## 环境假设
- 存在角色设定表（Character Sheet）、多角度参考图与风格参考集，带版本号。
- 存在 DCC 环境（Maya、Blender 一类）与渲染农场，支持批量渲染与序列导出。
- 存在骨骼生成、蒙皮、动画生成、插帧、超分、调色各阶段模型或工具。
- 存在镜头检查点与资产版本管理。
- 存在 scripts/ 目录承载确定性操作：抽帧、序列重命名、色彩校正与色彩空间转换、元数据写入、标识渲染、一致性比对、导出封装。
- 注意：目前**暂无专门的 AI 动画国家标准或行业标准**。质量阈值由项目自建，须明确标注为项目内控值。

## 上下文加载顺序（Context Budget）
1. 用户显式指令与本次镜头/序列的验收标准
2. 本文件（方向级）与组级 AGENTS.md
3. 角色设定表与多角度参考图（不可裁剪）
4. 风格参考集与色彩配置（不可裁剪）
5. 当前镜头在检查点中的上下文：前一镜头收尾姿态、运动基线、色调基线
6. 参考动作库与历史素材（可裁剪）
规则：角色设定表与风格参考集属于锚定上下文，任何情况下不得被裁剪。

## 工具契约
- 只读类（自由调用）：查询角色设定表、查询资产版本、查询检查点、检索动作库。
- 生成类（输出必须进校验关卡）：骨骼生成、蒙皮、动画生成、插帧、超分、调色。
- 校验类（阻断式）：一致性比对、滑步抖动检测、跨帧闪烁检测、标识校验、版权核查。
- 写操作类（二次确认 + 留痕）：覆盖资产版本、提交成片序列、提交渲染农场任务。
- 抽帧、色彩校正、元数据写入、标识渲染、一致性比对、导出封装一律走 scripts/。

## 任务执行流程（SOP）
1. 确认任务：项目、集数、镜头号、序列范围、输出规格（分辨率、帧率、色彩空间）。
2. 加载角色设定表、参考图与风格参考集，锁定版本号。
3. 骨骼生成：按角色类型（人形/四足/鸟类/怪兽）选择模型；检查物理辅助骨骼是否齐全。
4. 蒙皮：一键蒙皮后检查关节形变；复杂角色（3A 修型骨骼）自动化率较低，须人工复核。
5. 动画生成：多模态输入 + 姿态约束；输出精细至手指关节。
6. 插帧：用 MIB 类方法生成中间帧；检查滑步与抖动。
7. 两阶段渲染（若采用）：第一阶段保运动流畅，第二阶段保细节与跨帧稳定。
8. 超分与调色：提升至目标分辨率（如 4K）；统一色彩配置。
9. 一致性校验：面部/结构/语义一致性 + 帧间运动差异 + 滑步抖动率。
10. **人为参与节点**：在关键帧确认与细节渲染前设置强制人工确认，不得全部后置。
11. 标识处理：视频起始画面与播放周边添加显著提示标识，并写入隐式元数据。
12. 导出与回读：导出封装后回读，确认元数据与标识在位、色彩空间正确。
13. 更新检查点：记录版本、校验结果、失败原因、成本。

## 验证与证据要求
- 每个序列必须留存：候选路径、一致性比对分数、滑步抖动率、通过/不通过判定。
- 引用一致性阈值时必须标注为"项目内控值"或"方案建议值"，**不得表述为行业标准阈值**。
- 标识必须给出证据：起始画面标识位置 + 隐式元数据回读结果。
- 成本数据必须标注为"行业估算"或"项目实测"，不得混用。
- 引用效果数据必须标注口径层级。

## 失败与升级策略
- 角色漂移：回滚到最近检查点，重新加载设定表后重生成；连续 2 次失败升级人工。
- 滑步 / 抖动 / 扭曲：检查姿态约束强度与插帧参数；必要时回退到关键帧手工修正。
- 跨帧闪烁：检查超分与调色阶段的跨帧稳定性；启用特征传递机制。
- 蒙皮关节穿帮：复杂角色人工复核修型骨骼；自动化率不足属预期内，不得强推自动结果。
- 两阶段渲染破坏运动一致性：检查第二阶段是否改变了运动特征，必要时冻结第一阶段输出。
- 升级必须携带：项目、集数、镜头号、检查点记录、比对证据、已尝试处理。

## 安全与合规红线
- 视频起始画面及播放周边必须添加显著的提示标识（《标识办法》第四条第四项），并写入隐式元数据。
- 不得恶意删除、篡改、伪造、隐匿标识（第十条）；去标识日志留存不少于六个月（第九条）。
- 动画形象版权：不得生成受版权保护的既有动画形象。
- 训练素材与参考图必须有授权台账。
- 禁止使用未授权真人肖像作为角色参考。

## 禁止事项
- 禁止把项目内控阈值表述为行业标准阈值。
- 禁止跳过人为参与节点与一致性校验。
- 禁止把角色状态存放在模型会话里；必须写入检查点。
- 禁止让模型逐 token 生成抽帧结果、调色参数、元数据与导出配置。
- 禁止在无书面授权的情况下使用既有动画形象与真人肖像。
- 禁止编造标准编号与案例数值。

## 输出格式
项目 / 集数 / 镜头号 / 角色设定表版本 / 输出规格 / 各阶段产物路径 / 一致性比对结果 / 滑步抖动率 / 标识证据 / 人为参与节点记录 / 待签核项 / 责任人 / 检查点更新记录 / 遗留问题

## 评估与自检
- 角色设定表是否进入上下文且未被裁剪？
- 人为参与节点是否前置（而非全部放在最后）？
- 一致性阈值是否标注为项目内控值？
- 标识是否可回读？
- 本次失败样本是否已纳入评估集？

4.2 SKILL.md Specification

4.2.1. SKILL.md (AI Animation · Animation Shot Delivery with Character Consistency)

---
name: ai-animation-shot-consistency-delivery
description: AI 动画镜头交付技能。当需要从角色设定表出发生成一个动画镜头或序列，并完成骨骼与蒙皮检查、姿态约束、插帧与超分、角色一致性比对、滑步抖动检测、显著标识写入与镜头检查点更新时使用。适用于 3D 角色动画、AI 辅助二维动画与游戏过场动画场景。
version: 1.0
created: 2026-09-12
---

# AI 动画 · 角色一致性的动画镜头交付

## 适用场景
- 从角色设定表与分镜出发生成一个动画镜头或短序列。
- 需要在关键帧 → 中间帧 → 细节渲染 → 合成的分层流程中保持角色与动作一致性。
- 需要把每个镜头的生成状态与校验结果写入镜头检查点，支持单独重生成。

## 前置条件
- 已加载方向级 AGENTS.md 与组级 AGENTS.md。
- 角色设定表（Character Sheet）与多角度参考图可用且带版本号。
- 风格参考集与色彩配置可用且带版本号。
- 骨骼生成、蒙皮、动画生成、插帧、超分、调色工具可用。
- scripts/ 中存在抽帧、一致性比对、滑步抖动检测、元数据写入、标识渲染、导出封装脚本。
- 已确定项目内控的一致性阈值与滑步抖动率阈值，并经动画导演确认。
- 已确定人为参与节点（关键帧确认、细节渲染前）。

## 输入
| 输入项 | 说明 | 必需 |
|---|---|---|
| 角色设定表引用 | 角色 ID + 设定表版本 + 多角度参考图 | 是 |
| 分镜与表演描述 | 景别、运镜、表演要求、时长 | 是 |
| 输出规格 | 分辨率、帧率、色彩空间、是否 4K 超分 | 是 |
| 姿态约束 | 关键点来源与约束强度 | 否 |
| 一致性阈值 | 面部/结构/语义一致性与帧间运动差异的项目内控值 | 是 |
| 前一镜头检查点 | 收尾姿态、运动基线、色调基线 | 是 |

## 输出
- 动画镜头或序列（含候选）
- 一致性比对报告（面部/结构/语义 + 帧间运动差异）
- 滑步抖动检测报告
- 标识证据（起始画面标识位置 + 隐式元数据回读结果）
- 人为参与节点记录与镜头检查点更新记录

## 执行步骤
1. 从镜头检查点读取前一镜头收尾姿态、运动基线与色调基线。
2. 加载角色设定表、参考图与风格参考集，锁定版本号。
3. 骨骼与蒙皮：按角色类型选择模型；检查物理辅助骨骼；复杂角色人工复核修型骨骼。
4. 动画生成：多模态输入 + 姿态约束，输出精细至手指关节。
5. 插帧：生成中间帧，检查滑步与抖动。
6. 两阶段渲染（若采用）：先保运动流畅，再保细节与跨帧稳定。
7. 超分与调色：提升至目标分辨率，统一色彩配置。
8. 一致性比对：面部 / 结构 / 语义一致性 + 帧间运动差异；与项目内控阈值比对。
9. 滑步抖动检测：统计异常帧比例。
10. **人为参与节点**：关键帧确认与细节渲染前，由动画导演确认。
11. 标识处理：脚本在起始画面与播放周边渲染显著标识，写入隐式元数据。
12. 导出封装并回读校验（含色彩空间校验）。
13. 更新镜头检查点：版本、比对分数、失败原因、成本、责任人。

## 质量标准（DoD）
- 一致性比对全部通过，且留存比对分数；阈值明确标注为项目内控值。
- 滑步抖动率在阈值内。
- 起始画面与播放周边有显著标识；隐式元数据可回读；去标识日志留存不少于六个月。
- 人为参与节点已执行并留痕。
- 镜头检查点已更新，本镜头可单独重生成。
- 引用指标（如 ArcFace ≥ 0.6、DINO-I ≥ 0.45、CLIP-I ≥ 0.3、帧间运动差异 ≤ 5%）时标注"方案建议值，非标准阈值"。

## 常见失败与处理
- 角色漂移：设定表被裁剪 → 重新加载并固化为不可裁剪区。
- 滑步：脚部与地面相对位移失真 → 检查姿态约束强度与地面接触约束。
- 抖动：相邻帧无意义位移 → 检查插帧参数与超分阶段的跨帧稳定性。
- 扭曲：肢体非物理形变 → 回退到关键帧手工修正，不靠加大模型参数硬推。
- 跨帧闪烁：超分或调色引入 → 启用特征传递机制，冻结运动特征。
- 两阶段渲染破坏运动一致性：第二阶段改变了运动特征 → 冻结第一阶段输出作为约束。
- 蒙皮关节穿帮：复杂 3A 角色修型骨骼自动化率有限（约 50%+）属预期 → 人工复核，不强行自动。
- 异形角色骨架错误：确认模型支持范围（人形 / 四足 / 鸟类 / 怪兽）。

## 示例
任务：某项目第 3 集 SC-0315 镜头，角色 B 从静止到奔跑，时长 4 秒，输出 4K / 30 fps。
输入：角色 B 设定表 v2.4；多角度参考图；风格参考集 v1.8；前一镜头检查点；项目内控阈值。
执行：骨骼生成（含物理辅助骨骼）→ 蒙皮 → 姿态约束动画生成 → MIB 类插帧 → 两阶段渲染 → 4K 超分 → 一致性比对（帧间运动差异 3.1%，通过）→ 滑步检测（异常帧 0.4%，通过）→ 导演确认关键帧 → 标识渲染 → 导出回读 → 更新检查点。
输出：4K 序列 + 比对报告 + 滑步检测报告 + 标识证据 + 签核记录 + 检查点记录（SC-0315 = passed）。

4.3 Landing Checklist

5. Conclusion

The AI animation direction presents a sharp contrast: engineering practice leads, while standards construction is blank.

On the one hand, engineering practice has made substantial progress. Tencent VISVISE released a complete AI full-pipeline 3D character animation pipeline at SIGGRAPH Asia 2025; MotionGen Beta reduced the probability of anomalies such as foot sliding, jitter, and distortion by 6 times; MIB generates 200 frames in 4 seconds; overall production efficiency improved by more than 8x and has been applied to nearly 100 game projects. The Beijing Municipal Science and Technology Commission project compressed the production cycle of a single animation work from the typical 2~3 years to 5~6 months and supports 4K upscaling output.

On the other hand, no dedicated AI animation national standard or industry standard was found. This leaves consistency assessment without a unified metric: the currently circulating metrics such as ArcFace ≥ 0.6, DINO-I ≥ 0.45, CLIP-I ≥ 0.3, and inter-frame motion difference ≤ 5% are all proposed recommended values rather than standard thresholds. In engineering practice for this direction, "project-internal values" and "industry-standard values" must be clearly distinguished — this is not only a rigor requirement but also a necessary measure to prevent misuse downstream.

On the technical path, the two case studies give the same answer: consistency is achieved not by a stronger model but by state carriers and human-in-the-loop nodes. The Beijing Municipal Science and Technology Commission project integrated the "temporal memory module, cross-frame attention, and feature propagation mechanism" and explicitly added "human-in-the-loop nodes in the generation pipeline"; Tencent VISVISE uses an autoregressive framework with large-scale pretrained motion-transition representations to carry motion state. What they share is — translating the consistency problem from a model-capability problem into an engineering-structure problem.

Information Gap Statement

6. References

1. Identification Methods for AI-Generated and Synthetic Content — Cyberspace Administration of China, Ministry of Industry and Information Technology, Ministry of Public Security, National Radio and Television Administration, 2025. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
2. Building the Identification System Through Multiple Measures to Promote the Healthy Development of AI in the New Era — National Internet Emergency Center, 2025. https://www.cac.gov.cn/2025-09/06/c_1758880709361356.htm
3. Tencent Games VISVISE Releases the Industry's First AI Full-Pipeline 3D Animation Pipeline at SIGGRAPH Asia — ifeng.com, 2025. https://i.ifeng.com/c/8pBbzJLKCtV
4. Research, Development and Application of an AI Full-Pipeline Animation Production Platform (project number Z241100007624006) — Beijing Municipal Science and Technology Commission results database. https://www.ncsti.gov.cn/kcfw/zcg/kjcgxxxt/cgxq/?id=16253
5. Research Report on New Media Run by Mainstream Media (2024–2025, including CMG AI animation works) — People's Daily Online, 2025. https://sc.people.com.cn/BIG5/n2/2025/1030/c345167-41396739.html
6. Embracing the Intelligent Wave, Swimming Toward the Depths of Change — Xinhua News Agency, 2025. https://www.news.cn/20251114/1c01598d836c449bbfd54267b6ecea6d/c.html
7. Generative AI Opens a New Era for Media Development — Interpretation of the "Central Radio and Television Station AI Development White Paper (2025 Edition)" — Modern TV Technology, 2025. https://tvtech.cctv.com/tuijian20250707/index.html
8. AGENTS.md Official Site — Agentic AI Foundation (Linux Foundation). https://agents.md/
9. Lesson 15: AGENTS.md - giving agents project context — Addy Osmani, 2026. https://addyosmani.com/agents/15-agents-md/
10. Agent Skills Specification — agentskills.io. https://agentskills.io/specification
11. Equipping agents for the real world with Agent Skills — Anthropic, 2025 (updated 2025-12-18). https://claude.com/blog/equipping-agents-for-the-real-world-with-agent-skills
12. Building custom Skills for Claude — Anthropic Claude Cookbook, 2025. https://platform.claude.com/cookbook/skills-notebooks-03-skills-custom-development