Insights · AI and visual production
Changing a character’s outfit, transforming a sketch or animating an image can solve part of a production task. The difficulty comes when each correction also changes what was already approved. For a brand, the improvement worth examining is how much control it retains over the result.
In my video, I present the updates to ChatGPT Images 2.5 and 2.6. This article complements the demonstration with a communication and audiovisual production perspective: what to test, how to evaluate a deliverable and where professional intervention remains necessary. It does not attribute unmeasured savings or accuracy to those tests.
In the previous analysis of ChatGPT Images 2.5 for brands, I suggested evaluating the tool through cost per approved asset. That criterion remains useful with 2.6: receiving a response sooner does not prove a campaign finishes sooner. Selection, revisions and finishing are part of the work.
What changes from Images 2.5 to Images 2.6?
OpenAI introduced Images 2.5 on September 8, 2026: better reference fidelity and iterative editing, alongside Sketch, templates, comments and prompt sharing. Images 2.6 followed on November 4, with GIFs and improvements in speed and consistency. These are provider claims; their effect on a project needs testing.
The relevant test is preserving one approved decision while changing another. If the product, character or message changes without authorization, the asset needs correction, however attractive its finish.
From a sketch to a reference that supports decisions
A simple drawing can communicate a composition better than a lengthy paragraph. It can establish product placement, headline space and the relationship between elements. In a commercial proposal, this initial structure helps people discuss the intention before judging the finish.
Consider a restaurant campaign whose objective is to present a lunch menu and encourage reservations. A sketch can reserve the top area for the offer, put the dish in the foreground and leave a clear closing area for contact information. Evaluate the generated image against that purpose: if the background dominates and makes the offer harder to read, revise the composition.
The same logic applies to a storyboard. References support discussions about lighting, scale and framing before filming. However, a generated scene does not establish that a location exists or a camera movement is feasible. The team needs to turn the concept into a plan achievable with its resources.
Iterative editing requires a master reference
In professional production, every approval establishes a baseline. When changing a person’s surroundings, also review their face, clothing and proportions. When changing a product background, preserve its controls, connectors, packaging and apparent dimensions. Consistency requires comparison rather than reliance on an overall impression.
I suggest saving a master reference and a short list of elements that must remain unchanged. Then request one correction per round and compare versions. This makes it easier to identify when a change introduces an error and avoids accumulating edits on a baseline that has already lost fidelity.
There is a practical boundary: when an area must match the original exactly, preserve it through compositing. Official logos, regulated labels, prices and promotional terms should remain editable where possible. Regenerating them adds unnecessary verification and can alter the meaning of the asset.
GIFs: decide what moves and why
Movement can direct attention, explain a gesture or extend a scene. It can also distract. Before animating an asset, I would define a single focal point: steam from a drink, a character’s movement or a brief action that supports the message. Animating everything often weakens hierarchy.
For a technology product, for example, movement on a screen could draw attention but should not invent device capabilities. For an illustrated flower, animation can develop the character of the asset without compromising commercial information. Each application requires a different review.
Approval should cover the complete loop: transitions, distortions, elements appearing or disappearing and text stability. Then check file size, dimensions and compatibility at the actual destination. A GIF playing inside the chat is not yet an asset prepared for every website, advertisement or edit.
Reusable references and availability: two important distinctions
The release notes document saved references for reuse in other chats and GIF playback adjustments. They also state that GIF creation is rolling out and templates are unavailable in Work. Check the interface and limits of the account you will use.
For a team, retrieving a reference does not replace a production archive. I recommend recording its origin, usage authorization, purpose and approved version. A photograph may define atmosphere without authorizing product changes. That distinction should accompany the asset when another person takes over the project.
Also distinguish ChatGPT from integrations: the 2.6 announcement describes API features such as SVG export and complexity control. Do not assume these controls appear in every conversation. An integration requires assessment of its particular capabilities and costs.
What to test before adopting the tool for a campaign
These are evaluation proposals, not guaranteed outcomes. Define approval conditions before generating so visual appeal does not change the criteria midway through the process.
| Need | Work test | Approval criterion |
|---|---|---|
| Explore a campaign | Two visual directions from the same brief. | Each proposal communicates the message and is distinct without losing identity. |
| Place a product in a setting | Change the surroundings in an authorized photograph. | The device or packaging retains the original characteristics. |
| Plan a video | Create references for three consecutive shots. | Characters, objects and space maintain sufficient continuity to explain the sequence. |
| Animate an asset | Add localized movement to a short loop. | Movement supports the message and preserves critical elements. |
A method for measuring the value of the process
1. Define the brief and the exploration boundary
Write down the audience, objective, channel, format and main message. Add what can be reinterpreted and what must remain exact. Set an exploration time limit as well: generating alternatives without selection criteria can consume the supposed savings.
2. Review fidelity, communication and finish separately
First check whether the representation is correct. Next assess whether the message is understandable and the hierarchy works. Finally review technical quality, cropping and readability at publication size. An image can pass one of these criteria and fail the others.
3. Record work through to approval
Add preparation, generation, selection, correction and finishing time. Record discarded attempts too. Cost per approved asset combines tool consumption and professional time, divided by assets that meet the brief. Compare it with your usual process under equivalent conditions.
Example instruction for evaluating an edit
This prompt proposes a controlled test using an authorized image. Its value lies in enabling subsequent comparison, not guaranteeing every condition will be met:
Using the attached photograph, prepare a horizontal 16:9 concept for a corporate website. Preserve the product’s shape, proportions, controls and orientation. Change only the setting to a bright, restrained workspace. Maintain coherent lighting and leave clear space on the left for a headline. Do not add text, logos or accessories. If a product detail is unclear, do not fill it in with invented features.
After generation, compare the proposal with the original. If the product changes, decide whether to repeat the test or preserve the object through manual compositing. Continuing revisions indefinitely is not always the most efficient option.
Video: the updates in practice
The accompanying video presents the updates to Images 2.5 and 2.6. It will be available to watch when it goes public on YouTube.
Frequently asked questions about ChatGPT Images 2.5 and 2.6
Does better consistency guarantee an identical identity?
Do not assume so. My production criterion is to compare each deliverable with the approved reference. A general resemblance may be insufficient when representing a specific product or person.
Does faster generation reduce campaign costs?
That can only be established after measuring the complete process. If discarded attempts or finishing work increase, generation speed may not translate into savings.
What should remain outside generation?
Where accuracy is essential, I recommend retaining official logos, prices, commercial terms and other critical copy as editable elements. Final review should check both their content and presentation.
Does a GIF replace video production?
A short loop can address a specific visual need. A video involving narrative, voice-over, editing and pacing requires a production process suited to those objectives.
The tool adds value when it improves a decision
Practical progress means developing an understandable proposal, revising it with control and turning it into an asset that serves its purpose. The number of generated images says little about that ability. The useful benchmark is the quality of approved work and the effort needed to achieve it.
If you want to integrate AI into your company’s visual production, start with a concrete need and an evaluation criterion. Through strategic communication and audiovisual production, I can help you define that process. Explore my work and contact me at marcochile.com.
Sources and additional information