RemixFig is an open-source prompt workflow for researchers who generate beautiful framework diagrams with AI tools but need an editable version for their papers. No code, no installation, no API key required.
AI tools like ChatGPT and Gemini can generate stunning academic figures. But the output is always a flat PNG. You cannot edit the text, move the boxes, or swap the icons. Recreating the same figure from scratch loses all the visual quality.
RemixFig is a 5-step prompt sequence that instructs ChatGPT to analyze your figure, extract every element as structured data, rebuild the layout as a native editable slide, regenerate all icons at high quality, and assemble the final editable file. The result is a slide file where every box, text label, arrow, and icon is independently editable.
-
Step 1 sends the figure to ChatGPT and receives a detailed JSON describing every element in the diagram including boxes, text labels, arrows, and icons with percentage-based coordinates and style attributes.
-
Step 2 uses that JSON to generate an initial editable slide where all boxes, text, and arrows are native shapes. At this stage the icons are embedded as low-resolution screenshots and may not look sharp.
-
Step 3 asks GPT to regenerate every icon from the original figure as a clean high-quality vector-style version arranged on a blank canvas, along with a JSON mapping each icon to its original position.
-
Step 4 extracts precise coordinate and attribute data for each icon.
-
Step 5 replaces the low-resolution icon screenshots in the editable slide with the high-quality icons from Step 3, placed at their precise original positions.
💡 The final output may still require minor manual adjustments in the slide editor, such as repositioning a text box or resizing an element. The video ends by opening the output slide and clicking through individual elements to demonstrate that every box, text label, arrow, and icon is independently editable.
The demo video shows the complete 5-step process using a sample academic framework diagram as input.
👉🏻 Input AI-generated PNG
(Figure from: Chen, Y. (2026). Probabilistic Human Digital Twin (PHDT). Zenodo.)
👉🏻 Output English editable slide screenshot
Download English editable slide
👉🏻 Output Chinese editable slide screenshot
Download Chinese editable slide
Open a new ChatGPT conversation. Upload your figure as a PNG or JPG. Then send the 5 prompts below one at a time, waiting for GPT to fully respond before sending the next.
✨ English version: See English prompts
✨ Chinese version: See Chinese prompts
-
A ChatGPT Plus account with GPT-5.5 access is required. No other tools or setup needed. Use GPT-5.5. Earlier models produce inconsistent results.
-
Wait for each step to fully complete before sending the next prompt. The JSON outputs in Steps 1 and 4 will be long. This is normal. Just scroll past and continue.
-
The final slide may need minor text box adjustments in a few places. The overall layout and icon placement will be accurate.
-
Icons that are 3D renders or photographs rather than flat vector icons will be preserved as images and may look slightly softer than the original.
If RemixFig helps your research workflow, please consider citing or starring this repo.
@misc{chen2026remixfig,
author = {Yuyan Chen},
title = {RemixFig: A Prompt Workflow for Making AI-Generated Figures Editable},
year = {2026},
publisher = {GitHub},
url = {https://github.com/Yukyin/RemixFig}
}Noncommercial use is governed by LICENSE (PolyForm Noncommercial 1.0.0). Commercial use requires a separate agreement. See COMMERCIAL_LICENSE.md.
📨 Commercial inquiries: yolandachen0313@gmail.com