Sivia is a scientific-figure plugin for Codex and Claude Code, built for method illustrations in papers, lab meetings and thesis defenses. Give it a paper: Sivia maps the method and contribution, draws on an extensible external knowledge base, and generates an overview. After you approve the design and request the PowerPoint, it delivers an editable PPTX for PowerPoint or WPS.

Keep editing the result: change labels, move modules, adjust arrows and replace experiment images. Adapt the same method figure to a paper layout or a presentation.

| Capability | What you can do with it | | --- | --- | | Editable PowerPoint delivery | Work with native text, modules and arrows, with complex artwork and experiment images kept as separate assets. | | External knowledge base | Draw on examples matched by domain, scientific objects, method relationships and composition, including your own lab references. | | Local revisions and reusable experience | Revise the requested region while preserving approved layout choices; retain accepted cases and useful feedback for later work. |

Delivery: Paper → Overview preview → Approval and PowerPoint request → Editable PPTX + preview.

Sivia also reconstructs existing scientific figures and supports editable draw.io output. Its RSI research direction explores how drawing skills and the methods used to improve them can evolve over time.

Knowledge base

Connect your drawing workflow to an extensible external knowledge base. Sivia ships with curated scientific-figure examples and can also use reference figures, lab styles and approved work you supply. These references guide composition, hierarchy, spatial relationships and visual expression; the current paper determines the scientific content.

The library contains 10 approved cases across the five areas below, alongside 3 separately marked review drafts. The gallery shows overview images, with production records and source notes available for each case.

Browse the classified knowledge base · Extend the case library

3D reconstruction and dynamic geometry

Neuralangelo: High-Fidelity Neural Surface Reconstruction User-selected / active Full-size image · Production record · Case details	VGGT: Visual Geometry Grounded Transformer User-selected / active Full-size image · Production record · Case details
MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videos User-selected / active Full-size image · Production record · Case details	D4RT Review draft · not admitted Full-size image · Production record · Case details

Agents, tools and retrieval

ReAct: Synergizing Reasoning and Acting in Language Models User-selected / active Full-size image · Production record · Case details

AutoTool Review draft · not admitted Full-size image · Production record · Case details

Molecular modeling and AI for Science

DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking User-selected / active Full-size image · Production record · Case details

SigmaDock Review draft · not admitted Full-size image · Production record · Case details

World models and embodied perception

Learning View-invariant World Models for Visual Robotic Manipulation User-selected / active Full-size image · Production record · Case details	RoboSpatial: Teaching Spatial Understanding to 2D and 3D Vision-Language Models for Robotics User-selected / active Full-size image · Production record · Case details
Reconstructing People, Places, and Cameras User-selected / active Full-size image · Production record · Case details

Multimodal and motion-aware learning

Efficient Motion-Aware Video MLLM User-selected / active Full-size image · Production record · Case details

Dense-SfM: Structure from Motion with Dense Consistent Matching User-selected / active Full-size image · Production record · Case details

Quick start

After installation, attach your paper in Codex or Claude Code and describe the figure you need.

1. Upload a paper, generate an overview

Use Sivia to create a scientific overview of this paper.
Highlight its contribution, key modules, inputs and outputs in a compact composition for the paper.

Sivia handles paper understanding, reference selection, composition and image generation. Review the actual image and decide what to change.

2. Approve the design, generate an editable PowerPoint

I approve this overview. Use Sivia to make an editable PPTX with the same layout and content.
Use native objects for text, modules and arrows, and keep artwork and experiment images separate.

Receive PPTX + preview. Open it in PowerPoint or WPS to adjust wording, module positions and assets. If you already have an approved scientific figure, start here.

3. Make a targeted revision

Move the output module from the right to the bottom, reroute its arrows and enlarge the labels.
Preserve the other regions, content and colors. Update the editable PowerPoint and preview.

Bring your own knowledge base: attach reference figures or provide a case directory, then ask Sivia to use their composition and palette while designing around the current paper's method.

RSI

Recursive Self-Improvement · Research direction

A revision can contribute to the next drawing task. Sivia currently uses a design, drawing, review and correction workflow, and retains approved work, reference assets and useful feedback as local cases that can be reused in later tasks.

The RSI direction studies both drawing skills and the methods that update those skills: propose rules from revision feedback, evaluate them on new tasks, then explore recursive improvement of the updating strategy itself. The current source-available release provides editable figure production and case accumulation; recursive updater development is a research direction.

Installation

Codex

Run in a terminal, then start a new task:

codex plugin marketplace add exsinger-hub/Sivia --ref main
codex plugin add sivia@sivia

Claude Code

Run inside Claude Code:

/plugin marketplace add exsinger-hub/Sivia
/plugin install sivia@sivia

Development and documentation

Full workflow · Changelog · Contribute to the knowledge base

Run checks from the repository root:

node scripts/sync-plugin-metadata.mjs --check
node --test tests/*.test.mjs
python -m unittest discover -s tests -p "test_*.py"

License and commercial use

Non-commercial use is permitted. Commercial use requires prior written authorization from gatina.

Sivia-authored contributions are available under the Sivia Non-Commercial License 1.0. This includes personal learning, teaching and non-commercial academic research. Business use, paid services and commercial product integration require a commercial license request.

Upstream components retain their MIT license; earlier MIT releases keep their original terms. See third-party and historical notices.

Acknowledgments

Built on Scientific Illustrator by icebird1998（一个地质博士）, retaining its upstream MIT license and copyright. Sivia extends its editable drawing backends and Designer–Drawer–Reviewer–Corrector process with manuscript understanding, template-led ImageGen drafts and approval-locked reconstruction.

The bilingual README navigation and product-oriented organization take inspiration from CC Switch; its product claims and promotional content are not reused.

Thank you for using Sivia. Created by gatina.