保留真实像素Preserve real pixels
完整主体直接裁切；矩形框重叠但轮廓可分时，用前景掩码清理并重新排版，不把“重新生成”伪装成“提取”。Complete subjects remain source crops. Separable overlaps use foreground masks and clean re-layout, never disguising generation as extraction.
Design report / Release 0.1.0
把整张创意图，拆成真正可用的独立资产。 Turn one creative sheet into usable individual assets.
UltraSplitter 强大的智能拆件工具，可自动识别并批量提取画面中的角色、道具等主体元素，规范输出干净的独立资产素材，并支持残缺主体自动补全。SKILL安装，即可使用。 UltraSplitter is a powerful intelligent asset extraction tool that automatically identifies and batch-extracts characters, props, and other subjects, delivers clean and standardized individual assets, and supports automatic completion of incomplete subjects. Install the Skill to get started.
CASE 01 / VERIFIED RUN
INPUT SHEET
DELIVERED ASSETS
不均匀角色多视图
4 个原图裁切 · 未生成像素
v0.1.0当前版本Current release
Python ≥ 3.10运行环境Runtime
Schema v3可审计清单Auditable manifest
MIT开源许可Open-source license
01 / EXECUTIVE SUMMARY
不是均匀切格子，而是交付每一个可用主体。Not equal slices. Every usable subject, delivered.
位置不齐、比例混杂、轮廓相接、主体触边时，平均裁切会直接损坏结果。UltraSplitter 将语义判断与像素处理分开，让每次输出都有明确来源和状态。Equal slicing breaks when spacing, scale, silhouettes and edges stop behaving like a grid. UltraSplitter separates semantic judgment from pixel operations, giving every output a traceable origin and explicit state.
补全必须先批准Approve repair first
可识别但被截断或遮挡的主体会进入修复组；系统先生成请求包，不主动调用模型。每组最多接收两次结果。Recognizable clipped or occluded subjects enter repair groups. The system prepares a request without calling a model, with no more than two ingested attempts per group.
结果可检查、可追溯Reviewable and traceable
命名图片、紧凑联系表、坐标、路由依据、来源、审批和评估状态统一写入交付目录与 manifest。Named images, compact contact sheets, coordinates, route evidence, provenance, approvals and evaluations ship together with the manifest.
02 / VERIFIED SHOWCASES
三类真实输入，三种不同的拆分难题。Three real inputs. Three different splitting problems.
以下结果来自当前工作流。预览只负责让结果易读，不改变实际交付的独立文件。These results were processed by the current workflow. Previews improve readability without changing the delivered individual files.
03 / PROCESS MODEL
语义决定“是什么”，代码决定“像素在哪里”。Semantics decide what it is. Code decides where the pixels are.
Scan
从分隔线、边框、背景估计和连通域得到稳定候选。Find stable candidates from dividers, frames, background estimates and connected components.
Plan
多模态宿主组合区域，并记录完整性、身份置信和处理建议。The multimodal host groups regions and records completeness, identity confidence and treatment.
Route
在原图裁切、原像素重排与生成式重建之间选择明确路径。Choose explicitly between source crop, source composite and generated reconstruction.
Repair
只为可识别、满足门槛并得到批准的主体准备补全任务。Prepare completion only for recognizable subjects that pass the gates and receive approval.
Evaluate
检查数量、重复、背景、分辨率、触边、留白和身份一致性。Check count, duplicates, background, resolution, edge contact, margins and identity.
04 / OPERATING BOUNDARY
能力边界写清楚，才不会把“看起来成功”当成交付成功。Clear boundaries keep visual success from masquerading as delivery success.
当前适合Works well today
- 有分栏或边框的面板式图片Framed or divided panel layouts
- 透明或近似纯色背景中的分离主体Separated objects on transparent or near-uniform backgrounds
- 需要命名资产、联系表和可审计 manifest 的工作流Workflows needing named assets, contact sheets and auditable manifests
当前不承诺Not promised today
- 复杂语义实例分割Complex semantic instance segmentation
- 把生成式补全描述为恢复出的原始像素Describing generated completion as recovered source truth
- 对密集、触碰、透明或截断情况静默宣称成功Silently claiming success for dense, touching, transparent or clipped cases
05 / QUICKSTART
从一张图开始。Start with one image.
通过 Agent Skill 使用完整工作流，或直接运行 Python CLI。核心包不保存供应商密钥，也不主动联网。Use the Agent Skill for the full workflow or run the Python CLI directly. The core package stores no provider credentials and makes no network calls.
AGENT SKILL
npx skills@latest add \
https://github.com/PlevanTem/UltraSplitter/tree/main/.agents/skills/splitting-image-grids-by-contentPYTHON CLI
git clone https://github.com/PlevanTem/UltraSplitter.git
cd UltraSplitter
python -m pip install -e .
ultrasplit run input.png --name character-views