ChatGPT Images 2.5 brings sketch-to-image, steadier multi-turn edits, and reference-photo blending. We tested each feature hands-on to see what holds up.
What Is ChatGPT Images 2.5 And What Does It Change?
ChatGPT Images 2.5 is OpenAI's updated image generation model inside ChatGPT, released in September 2026 as the successor to GPT Images 2. The upgrade focuses on three areas: sharper detail, faster generation, and more precise multi-turn editing, plus new workflows like sketch input and an updated editing panel. You can find the model discussed in OpenAI's ChatGPT release notes when features roll out to your account.
Access is rolling out gradually. In the ChatGPT web app, you know you have Images 2.5 when the Images tab shows the updated icon. The sketch feature, editing panel, and templates appear under the plus menu in a new chat. The presenter, Bart Slodyczka, tested each feature live in his walkthrough, and this article follows the same order: sketch, precision edits, reference photos, and site mockups.
The most practical change is edit stability. Earlier versions tended to drift: after several iterations, faces changed, objects moved, and lighting shifted. In the tests below, stacked edits preserved nearly everything, which is what makes the model usable for real design work rather than one-shot generation.
How Does Sketch-To-Image Work In ChatGPT Images 2.5?
Sketch-to-image in ChatGPT Images 2.5 turns a rough drawing into a finished render by treating your sketch as spatial instructions rather than art. You open it from the plus menu, pick "sketch", and get a canvas with shapes and a freehand mode. Draw a layout, add a text description of each element, and the model builds an image that keeps your arrangement.
Example: Sketching A YouTube Studio
In the test, the presenter sketched a YouTube studio plan: a large square for the floor, a green rectangle for the desk, yellow circles for two camera positions, a purple rectangle for his seat, and a corner seat. He submitted the sketch, then added a text description explaining each element and asked for an aerial view of the room.
Generation took about 90 seconds. The result was a detailed 3D-style room render that kept the layout: desk, two monitors, both cameras, and both seats. Not everything matched perfectly, like camera angles being sideways, but overall consistency with the sketch was strong. He then edited the render, asking for a tall Monstera plant in a circled spot, and the plant appeared exactly where marked with almost no other changes to the image.
Sketch Elements And What They Became
| Sketch element | Intended meaning | Result in the render |
|---|---|---|
| Large square | Floor | Full 3D-style room base |
| Green rectangle | Desk | Desk with two monitors |
| Yellow circles (two) | Camera positions | Both cameras placed as marked |
| Purple rectangle + corner seat | Seating | Both seats included in the layout |
Editing And Collaboration
The updated editing panel lets you annotate, comment, remove backgrounds, erase, and resize within the image workspace. You can also share images or the prompt template directly, which matters if you work with a designer or a team.
How Consistent Are Multi-Turn Edits Compared To GPT Images 2?
Multi-turn edit consistency is the core upgrade of ChatGPT Images 2.5. To test it, the presenter made a series of small edits to the studio render: add a plant, remove a wall sign, swap the carpet for a kids' play rug. Each time, he compared before and after in the editing panel.
Results held up well. Removing the wall sign erased it completely with roughly a 1% lighting shift and nothing else changed. The carpet swap produced exactly the rug described while keeping the Airtable plant in identical positions. The main drift across edits was a slight darkening of color over several passes, which is minor but worth watching if color accuracy matters to your project.
The old model failed at this. With GPT Images 2, the presenter's experience making YouTube thumbnails was that after three or four generations, his face and appearance distorted or changed into a different person entirely. Objects like buildings also shifted position across iterations. Images 2.5 is much more stable, though not perfect: in the longer reference-photo test below, edited regions did still shift between runs.
Edit speed matched initial generation speed, around 90 seconds per edit, with progress stalling slightly in the last 10% of the progress bar.
Can It Combine Multiple Reference Photos Into One Scene?
Yes. Images 2.5 can take several reference photos and merge them into a single new scene, blending the subjects' appearances with the style and setting of another image. It handles composition, lighting, and detail preservation well, though stacked follow-up edits can drift slightly from earlier instructions over multiple passes.
OpenAI's own examples include turning a child's photo into a formal portrait, putting a costume on a dog while preserving every shadow, and combining three separate portraits of different people into one candid party photo.
To test the composition case, the presenter uploaded photos of three comedians he likes: Will Ferrell, Andrzej Grabowski, and Eddie Izzard. He also uploaded a sample party photo as a style reference and asked the model to put the first three people into one image matching the fourth, with a vintage look, glossy photo paper texture, and everyone embracing.
The result combined all inputs: the three faces, the red-toned wall, the hallway floor from the reference, and even a poster from the original background. Details like hands resting on a stomach and shadows matched closely. He then stacked more edits on top:
- Enlarge the red cups by 25%.
- Hang a skateboard horizontally showing its underside.
- Add a strawberry iron-on transfer on the middle person's shirt.
Here the limits showed. The cups grew larger than requested on the second pass and kept growing on later edits, and the whole image shifted slightly rightward each time. The model appeared to reprocess the earlier "bigger cups" instruction along with new ones. Still, after four stacked edits, the core scene, three people embracing with red cups and a skateboard on the wall, remained intact and photographic.
| Aspect | Result |
|---|---|
| Combining three faces into one scene | Accurate, all three preserved |
| Style transfer from party photo | Wall, floor, and poster carried over |
| Four stacked edits | Scene intact, but cups oversized and image drifted right |
| Face consistency vs previous model | Held across iterations instead of producing a different person |
With the previous model, the presenter said, four or five iterations on a face usually produced a completely different person.
What Are The Templates, And Can It Build A Real Website?
Images 2.5 ships with templates, including a site mockup template that turns a product image into a full website design — and it can go one step further, converting that mockup into a working single-file HTML landing page. The workflow below shows how a single product photo becomes a product sheet, a site mockup, and finally a real page, along with the time each stage took.
The presenter first asked the model to invent a fake product, a water bottle called Bart's Water with healthy orange and green colors, plus a product sheet showing the bottle from several angles, a logo close-up, and a cap close-up. That took 1 minute 42 seconds and produced a polished marketing sheet with tagline, bottle specs, and icon details.
He then downloaded that image, opened a new chat, applied the site mockup template, and uploaded the product sheet. The model asked about site purpose, target audience, and needed content; he told it to figure those out itself. After 4 minutes 15 seconds, it produced a complete mockup with the brand name at top, headline, icons, and layout, all derived from the product sheet's text and imagery.
The final test: convert the mockup into a real, single-file HTML landing page. This ran in pro mode and took 17 minutes 14 seconds. The output was a working page with hero section, an add-to-bag flow leading to a checkout page, and product imagery it generated itself along the way. Checkout was not connected, which the model acknowledged. For a business owner with one product photo, the workflow of generating five product angles and wrapping them in a site is now realistic in a single session, though complex effects like 3D scroll animations still need video frames the model cannot efficiently produce yet.
| Stage | Input | Output | Time |
|---|---|---|---|
| Product sheet | Fake product brief (Bart's Water) | Marketing sheet with angles, tagline, specs | 1 min 42 sec |
| Site mockup | Product sheet via site mockup template | Full mockup with brand, headline, layout | 4 min 15 sec |
| Real website | Mockup, run in pro mode | Single-file HTML page with add-to-bag and checkout flow | 17 min 14 sec |
FAQ
- When did ChatGPT Images 2.5 come out? It launched in September 2026, succeeding GPT Images 2. The presenter covered it on September 9, 2026, and features were still rolling out to all accounts at that time.
- How do I know if I have access to Images 2.5? Open the ChatGPT web app and check the Images tab. If you see the updated icon, you have the new model. The sketch feature appears under the plus menu alongside other new options.
- How long does one image take to generate? Most generations and edits in testing ran about 90 seconds, with the last 10% of the progress bar moving slower. A full site mockup took about 4 minutes, and converting a mockup to a real HTML page in pro mode took over 17 minutes.
- Does editing an image change the rest of it? Mostly no, which is the main upgrade. In testing, consecutive edits preserved roughly 99% of the image, with slight color darkening over many passes. Larger or repeated edits can still shift cropped regions and reapply earlier instructions.
- Can it combine photos of several people into one image? Yes. The presenter merged three separate portraits into one convincing party photo, keeping faces, shadows, and even background details from a reference image. Identity held up better than in previous models across later edits.
Turn Your Own Video Walkthroughs Into Written Guides
This article started as a 22-minute video, and every test above, the sketch, the stacked edits, the party photo, the 17-minute site build, survived the trip to text because the demonstrations were described clearly enough to verify. If you make videos like this, your best material is probably sitting unwatched in your archive as footage.
Skala Blog turns a YouTube video into a structured, searchable article. Paste your video URL, get a transcription, and generate a written piece your audience can skim, quote, and find through search instead of scrubbing a timeline. It keeps the demonstration's detail and gives it a second life on the page.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits