Hello OpenAI Product Team,
I would like to propose a high-impact capability for a future version of ChatGPT and ChatGPT Work:
Visual-to-Editable Source Reconstruction.
The core idea is simple:
A user uploads a single reference image, and ChatGPT reconstructs that image as a genuinely editable design file with separated, semantically meaningful layers.
Instead of merely generating a visually similar image, ChatGPT would reverse-engineer the visual structure and produce files such as:
PSD
AI
Layered PDF
SVG
Figma-compatible assets
Other structured design formats where technically appropriate
The key requirement is genuine editability.
For example, if a user uploads a poster, advertisement, presentation graphic, UI mockup, infographic, social-media design, package design, or other flat image, ChatGPT should be able to identify and reconstruct:
Background
Foreground objects
Photographic elements
Logos
Icons
Illustrations
Text
Typography
Shapes
Lines
Gradients
Shadows
Borders
Masks
Transparency
Adjustment layers
Effects
Reusable components
Vector elements
Raster elements
Each meaningful visual component should become an independently editable layer or object.
The result should not simply be a flattened image placed inside a PSD or PDF.
It should be a reconstructed source file.
1. Editable text should remain editable text
This is one of the most important requirements.
When ChatGPT detects typography, it should reconstruct the text as actual text layers whenever possible rather than converting everything into pixels.
The system could identify:
Text content
Font family or closest available equivalent
Font size
Weight
Tracking
Kerning
Line spacing
Alignment
Color
Stroke
Shadow
Transform
Rotation
Text box dimensions
If the exact font cannot be identified or legally provided, ChatGPT should clearly state that and suggest the closest compatible substitute.
The user could then simply click the text and rewrite it.
2. Shapes and graphics should become vectors whenever possible
Logos, icons, geometric shapes, lines, charts, diagrams, interface components, and other suitable elements should be reconstructed as vector objects rather than unnecessarily rasterized.
For Illustrator-compatible output, ChatGPT could recreate:
Paths
Groups
Compound paths
Clipping masks
Gradients
Strokes
Fills
Transparency
Text objects
This would make the output genuinely useful for professional design work.
3. Photographic objects should be intelligently separated
When a reference image contains people, products, furniture, scenery, or other photographic subjects, ChatGPT could use semantic segmentation and generative reconstruction to separate them into individual raster layers.
For example:
Layer 01 — Background
Layer 02 — Person
Layer 03 — Product
Layer 04 — Foreground object
Layer 05 — Shadow
Layer 06 — Logo
Layer 07 — Headline
Layer 08 — Subtitle
Layer 09 — CTA button
The user could then independently move, resize, recolor, replace, hide, or edit each element.
4. Reconstruct hidden areas when objects overlap
A particularly powerful capability would be intelligent layer completion.
Suppose text partially covers a photograph, or one object overlaps another.
Simply separating the visible pixels would leave holes behind.
ChatGPT could reconstruct the hidden portions of underlying layers using generative image completion so that each separated layer remains independently usable.
For example:
Original flattened image:
Person in front of background.
Reconstructed source:
Complete background layer
Complete isolated person layer
Removing the person would reveal a plausible reconstructed background instead of an empty silhouette.
This would dramatically improve practical editability.
5. Preserve visual fidelity
The user should be able to choose different reconstruction priorities.
Possible modes:
Maximum Visual Fidelity
Maximum Editability
Balanced Reconstruction
Vector-First Reconstruction
Print Production Mode
Web/UI Reconstruction Mode
For professional workflows, the system should preserve:
Canvas dimensions
Aspect ratio
Element positioning
Relative scale
Margins
Alignment
Color relationships
Typography hierarchy
Composition
Layer order
Where the exact original information cannot be recovered from a flattened image, ChatGPT should distinguish between:
Directly recovered information
High-confidence reconstruction
Approximate reconstruction
Generated missing information
This is important because AI should not pretend that information hidden in a flattened image has been perfectly recovered when it has actually been inferred.
6. Provide a reconstruction confidence report
Professional users would benefit from a short reconstruction report.
For example:
Text detection confidence: 98%
Font identification confidence: 82%
Object segmentation confidence: 96%
Vector reconstruction confidence: 91%
Background reconstruction: AI-generated
Original font unavailable: substitute used
This would make the system trustworthy rather than opaque.
7. Let users refine the reconstruction conversationally
After reconstruction, the user should be able to say:
“Separate the headline and subtitle into different text layers.”
“Convert this logo into vector paths.”
“Make the background independent from the person.”
“Keep the exact composition but replace the product image.”
“Convert all icons into vectors.”
“Group all footer elements.”
“Rename the layers professionally.”
“Make this Photoshop-ready.”
ChatGPT Work could then modify the actual source file rather than merely generating another flattened image.
This would create a true conversational design workflow.
8. Automatically organize professional layer structures
The system could create logical layer groups such as:
00_GUIDES
01_BACKGROUND
02_IMAGES
03_OBJECTS
04_GRAPHICS
05_LOGOS
06_TEXT
07_EFFECTS
08_ADJUSTMENTS
09_EXPORT
Layer names could automatically reflect their content.
For example:
BG_Main
Photo_Person
Product_Main
Logo_Brand
Headline
Body_Copy
CTA_Button
Shadow_Product
This would make AI-generated files significantly easier for human designers to continue editing.
9. Support PSD, Illustrator, PDF, SVG, and design ecosystems
The long-term value would increase substantially if ChatGPT Work could export reconstructed designs into multiple professional ecosystems.
Potential outputs could include:
Adobe Photoshop PSD
Adobe Illustrator AI
Layered PDF
SVG
Figma-compatible structured design
Canva-compatible design
PowerPoint objects where appropriate
Rather than forcing users into one application, ChatGPT could act as a universal design reconstruction layer.
10. Introduce an “Open as Editable Design” command
The user experience could be extremely simple.
A user uploads an image and selects:
Open as Editable Design
ChatGPT then asks:
Photoshop
Illustrator
Figma
Layered PDF
SVG
Other
The system analyzes the reference image, builds its internal scene representation, and generates the appropriate editable source file.
A more advanced command might be:
“Reconstruct this image as closely as possible as a fully editable Photoshop file. Keep all text editable, separate photographic subjects into layers, convert geometric elements to vectors where possible, recreate hidden background regions, and preserve the original layout.”
This would make sophisticated reverse-design workflows accessible through natural language.
11. Why this matters
Today, enormous amounts of visual content exist only as flattened images.
Users frequently encounter situations where:
The original PSD was lost.
The original designer is unavailable.
Only a JPEG or PNG remains.
A client wants to modify an old advertisement.
A company wants to localize an existing campaign.
A designer wants to study or rebuild a reference composition.
A presentation graphic needs to become editable.
A UI screenshot needs to become a design prototype.
An AI-generated image needs to become a production-ready design asset.
Currently, recreating these assets manually can require substantial professional labor.
AI image generation solves a different problem.
It generates new pixels.
What professional designers often need is something fundamentally different:
Recovering editable structure from pixels.
This could become one of the most valuable distinctions between ordinary AI image generation and AI-assisted professional design production.
12. Strategic opportunity for ChatGPT Work
ChatGPT already combines:
Vision understanding
Image generation
Image editing
Document generation
Coding
Computer interaction
File manipulation
Application integrations
Long-running workflows
Agentic reasoning
A Visual-to-Editable Source capability would connect these technologies into a powerful professional design workflow.
ChatGPT would no longer only answer questions about an image or modify the image.
It could understand the image as a design system.
The transformation would be:
Image
→ Visual understanding
→ Semantic scene decomposition
→ Layer reconstruction
→ Typography reconstruction
→ Vector reconstruction
→ Missing-region reconstruction
→ Editable source file
→ Human refinement
This would turn ChatGPT from an image-generation assistant into a genuine AI production partner for designers.
13. Integration with ChatGPT Work would make it substantially more powerful
ChatGPT Work could potentially go even further.
For example:
A designer uploads 30 flattened campaign graphics.
ChatGPT Work reconstructs each one as an editable source file.
It identifies shared components across the campaign.
It creates reusable logos, typography styles, brand colors, buttons, and graphic components.
It organizes the files.
It identifies inconsistent layouts.
It produces localized versions.
It exports final assets.
A workflow that might normally require many hours of repetitive design-production work could become a supervised AI workflow.
14. Professional-quality verification should be built in
Before exporting the reconstructed design, ChatGPT could automatically compare the reconstruction against the original reference image.
It could evaluate:
Pixel-level visual difference
Element positions
Typography placement
Color difference
Object boundaries
Layer completeness
Alignment
Missing elements
If the reconstruction is not sufficiently close, the system could automatically iterate before producing the final file.
The goal should be:
Reference Image
→ Reconstruction
→ Render
→ Compare
→ Correct
→ Re-render
→ Validate
→ Export
This verification loop would be particularly valuable for professional designers.
15. Copyright and provenance safeguards should be included
The feature should also provide appropriate controls for copyrighted materials.
Possible safeguards could include:
Clear provenance metadata
Warnings when reconstruction involves protected brand assets
Restrictions where legally necessary
User confirmation of rights for certain commercial uses
Identification of AI-reconstructed versus directly extracted components
These protections could coexist with legitimate workflows such as restoring one’s own lost design files, modifying authorized client materials, reconstructing internal company assets, education, accessibility, localization, and design-production work.
16. The larger opportunity
Generative AI has already made it possible to create sophisticated visual images from language.
The next major step should be making visual AI output structurally editable.
Pixels should not be the end of the workflow.
They should be one representation of a deeper editable design structure.
In the future, users should be able to move seamlessly between:
Prompt
Image
Layers
Vectors
Text
3D
Documents
Presentations
Code
Interactive experiences
ChatGPT could become the system that understands and transforms between all of these representations.
I believe Visual-to-Editable Source Reconstruction could become one of the most valuable professional capabilities in ChatGPT and ChatGPT Work, particularly for graphic designers, creative agencies, marketing teams, presentation designers, UI/UX designers, publishers, content creators, and enterprises managing large amounts of visual content.
The ideal experience would be remarkably simple:
Upload any design reference.
Say:
“Make this fully editable.”
And receive a properly reconstructed professional source file.
Not a screenshot inside a PSD.
Not a flattened PDF.
Not a visually similar imitation.
A genuinely editable design.
That distinction is critical.
If OpenAI can make visual content not only generatable but structurally editable, ChatGPT could fundamentally change how professional visual production is done.
Thank you for considering this proposal.