ChatGPT and Gemini Image Generation in One Shared Conversation

AI image generation is moving beyond one-shot prompts

The first wave of text-to-image tools was largely transactional: write a prompt, receive an image, change the prompt, generate another image.

That model is changing.

OpenAI describes modern ChatGPT image generation as capable of building on the text and images already present in a conversation. Its current image system also emphasises stronger consistency across multiple editing turns, where later edits can build on earlier work instead of starting from zero.

Google has taken a similar conversational direction with Gemini image generation and editing. Its image tools support targeted changes, image blending and multi-turn editing through ordinary language.

In other words, the unit of work is becoming the visual conversation, not just the prompt.

That matters because creative work is rarely finished in one instruction.

What is multi-turn image editing?

Multi-turn image editing means refining an image through a sequence of conversational instructions.

A simple workflow might look like this:

1. Generate a portrait of a hiker standing in the mountains.
2. Change the scene from daytime to night.
3. Keep the same composition but add moonlight.
4. Change the jacket colour.
5. Return to the earlier version and try a different atmosphere.

The value is continuity. You should not have to reconstruct the entire scene description every time you want one change.

Google explicitly calls this multi-turn image editing. OpenAI similarly describes image editing across longer conversations where each change can build on the work already done.

The concept becomes more interesting when more than one AI model can participate in that visual history.

ChatGPT image generation and Gemini image generation do not have to live in separate tabs

A common way to compare AI image generators is to give ChatGPT and Gemini the same prompt in two separate products.

That works, but it creates the same problem that appears in text-based multi-model work: fragmented context.

One tab contains the ChatGPT image. Another contains the Gemini image. A third conversation may contain your critique. If you want one model to work from an image made by another, you become responsible for moving files between them.

In Plurilog, both image generators can participate in the same discussion.

Ask for an image once and ChatGPT and Gemini can produce different interpretations of the prompt. Those outputs can then become visual evidence available to the panel instead of being treated as unrelated files.

This is the same design principle behind our broader approach to putting ChatGPT, Claude and Gemini in one shared AI conversation: the conversation should remain the source of truth.

ChatGPT vs Gemini image generation: compare the actual outputs

Search for “ChatGPT vs Gemini image generation” and you will find plenty of winner-and-loser comparisons.

The problem is that image generation is unusually prompt-dependent.

One model may produce the composition you prefer for a photoreal portrait. Another may follow a detailed layout instruction more closely. A model that performs well on one prompt may be less convincing on the next. Editing quality can also be a different question from first-pass generation quality.

So instead of assuming there is one universally best AI image generator, a more useful workflow is:

• send the same brief to more than one image model;
• keep the outputs in the same discussion;
• inspect the actual pixels rather than relying on descriptions;
• compare composition, lighting, realism, detail and instruction-following;
• choose the version that fits the task;
• continue editing the strongest candidate.

It is the visual equivalent of asking multiple AI models for different perspectives on the same problem.

If you are also deciding which text model to use, our ChatGPT vs Claude vs Gemini comparison explains why the answer often depends on the task rather than the logo.

What shared visual context actually means

“Shared multimodal context” sounds technical, but the idea is simple.

Text is not the only thing a conversation can remember.

A multimodal discussion can also contain photographs, screenshots, generated images, edited versions, PDFs and other visual material.

Shared visual context means the system keeps track of which of those images are relevant to the current discussion so a model can receive the right visual evidence when it needs it.

In a multi-model system, that becomes especially important. ChatGPT, Gemini and Claude are separate models from separate providers. They do not naturally share one native image history.

Plurilog therefore treats the visual discussion as its own persistent context. When a model needs an earlier image, Plurilog can recover the relevant visual source and attach the actual image to that model's turn.

The distinction between remembering that an image existed and actually giving the model the image again is crucial.

Why visual memory has to be more than text memory

Imagine an earlier image showed a woman wearing a teal jacket on a mountain ridge.

A text summary might say exactly that.

But if you later ask:

“Which version has better rim lighting around the subject's hair?”

the model needs the pixels.

A remembered description cannot substitute for inspecting the image itself.

That is why Plurilog's visual context can reopen historical image evidence for the model that needs it. When several images are being compared, the relevant set can be delivered together.

It also reduces a subtle failure mode in multimodal AI: a model confidently describing an image from conversational memory even though the image is not actually attached to its current turn.

Can ChatGPT edit an image generated by Gemini?

In Plurilog, yes.

If Gemini generates an image in the discussion, that image can remain part of the shared visual context. You can then address ChatGPT and ask it to edit the Gemini result.

For example:

That is not simply three AI answers next to one another. It is one visual object moving through a multi-model conversation.

The original image is preserved rather than destructively replaced, which makes it possible to compare the source with the edited version later.

Claude can still participate even when it is not the image generator

Image generation is only one role in a visual workflow.

In the current Plurilog panel, Claude does not generate images. But it can inspect images available to the conversation and contribute as a visual reviewer.

That means Claude can compare a ChatGPT generation with a Gemini generation, inspect an edited version, describe structural differences or discuss which output better satisfies a creative brief.

This matters because a multi-model visual workflow does not require every model to perform the same action.

One can generate. Another can edit. Another can critique. The useful part is that they can work from the same underlying visual material.

A practical shared-context image workflow

Here is what an end-to-end image conversation can look like:

1. Generate: ask ChatGPT and Gemini for an image from the same brief.
2. Compare: ask Claude, ChatGPT or Gemini to inspect the outputs and explain the differences.
3. Choose: identify the stronger composition or the version that better matches your goal.
4. Edit: ask an eligible image model to modify the selected version.
5. Iterate: continue with another edit without losing the working visual thread.
6. Reopen: later, ask the panel to compare the original, another model's version and the edited result.

The benefit is not simply fewer clicks. It is fewer opportunities for the context to fragment.

Where shared-context AI image generation is useful

Multi-agent image generation is becoming a research problem too

The idea of several AI agents participating in image generation and editing is also appearing in current research.

The 2026 AAAI paper Talk2Image describes a multi-agent system for multi-turn image generation and editing. Its authors argue that single-turn workflows struggle with iterative creative tasks and use dialogue history, specialised agents and feedback-driven refinement to improve control and coherence.

More recent work such as WeAgent-MMGenEdit explores persistent multimodal evidence, visual verification and agentic image generation and editing.

These systems are not the same as Plurilog, but they point in a similar direction: generative imagery is increasingly being treated as something that lives inside an ongoing reasoning process rather than as a one-off media endpoint.

Shared context does not make AI images automatically correct

More context and more models do not remove the limitations of generative AI.

Image models can still:

• misread part of a prompt;
• change details you wanted preserved;
• invent visual information;
• produce unrealistic anatomy, lighting or text;
• make an edit that is attractive but inconsistent with the source.

Another AI model can help you notice those issues, but agreement between models is not independent verification.

That is the same reason we recommend treating multiple AI perspectives as a way to expose differences, not as a majority-vote truth system. Our article on AI hallucinations and cross-checking AI answers covers that distinction in more detail.

The bigger shift: images are becoming part of the conversation

AI image generation started with a simple interaction:

prompt → image

Conversational editing changed that into:

prompt → image → edit → edit → refinement

Shared visual context adds another layer:

generate → compare → discuss → edit across models → reopen → compare again

That is the direction we find more interesting.

The image is no longer just an answer attached to one model's message. It becomes a persistent object inside the discussion.