Image generation by Gemini Flash 2.5
4 views
Skip to first unread message
Robert Brown (RLBrown)
Oct 17, 2025, 12:36:13 AMOct 17
to sv...@googlegroups.com
Hello SVMUG members and friends,
It has been about 8 months since my last presentation on AI image generation. There has been a lot of progress since then. In particular, the release of Gemini Flash 2.5 (which has the street name of "Nano Banana") has generated a lot of excitement. Here is a quick update.
Gemini Flash 2.5 is accessible via the web interface to Google tools, or as a third-party model under Adobe Photoshop generative fill or Adobe Firefly text-to-image tool. The public response was initially overwhelmingly positive, primarily driven by its viral success on crowdsourced AI testing platforms.
Core Strengths: Users and developers praised the model for solving critical weaknesses found in previous AI image models:
Character Consistency: The ability to maintain the identity, face, and style of a character or subject across multiple images and edits, which is crucial for storytelling and branding.
Conversational Editing: Its superior multi-turn editing capabilities allow users to converse with the AI to refine an image using simple, natural language prompts (e.g., "make it red," "add a hat") rather than requiring complex manual masking or software skills.
Speed: The "Flash" model was noted for its low latency and rapid creative workflows, enabling quick iterations.
But there are some problems. Requests for minor edits, e.g., "Change the color of ... ", tend to be ignored. The AI returns the unmodified image.
The AI also heavily restricts the kinds of images it can generate or modify. It won't work on depictions of historical figures and stays conservative about adult content.
To be consistent with the images I have shown you before, I used a variation of my standard prompt: Generate an image of a group of physicists writing and discussing equations on a whiteboard. The equations should relate to quantum field theory. The image should be of photographic quality. The group of physicists should include both genders and show diversity.
In the past, you may recall that the equations were gibberish, the people tended to have messed-up hands, and there were weird things like using felt-tip markers on a chalkboard. Arms would have no hands.
In the image I have attached below, the generation is much better. There are many flaws, but it shows a plausible understanding of the prompt. The equations are still gibberish, but now the symbols and form are mostly correct, just not intelligible. There are spelling errors. The girl in the center has a broken finger on her right hand. The girl on the top left has a marker with a red end cap but a green tip. The guy on the right has a blue marker that writes in black.
On the whole, a credible job. Do you see any other errors or flaws?
I believe Adobe made a wise decision by partnering with Google to integrate Flash 2.5 into Photoshop and Firefly.
-- Robert Brown
It has been about 8 months since my last presentation on AI image generation. There has been a lot of progress since then. In particular, the release of Gemini Flash 2.5 (which has the street name of "Nano Banana") has generated a lot of excitement. Here is a quick update.
Gemini Flash 2.5 is accessible via the web interface to Google tools, or as a third-party model under Adobe Photoshop generative fill or Adobe Firefly text-to-image tool. The public response was initially overwhelmingly positive, primarily driven by its viral success on crowdsourced AI testing platforms.
Core Strengths: Users and developers praised the model for solving critical weaknesses found in previous AI image models:
Character Consistency: The ability to maintain the identity, face, and style of a character or subject across multiple images and edits, which is crucial for storytelling and branding.
Conversational Editing: Its superior multi-turn editing capabilities allow users to converse with the AI to refine an image using simple, natural language prompts (e.g., "make it red," "add a hat") rather than requiring complex manual masking or software skills.
Speed: The "Flash" model was noted for its low latency and rapid creative workflows, enabling quick iterations.
But there are some problems. Requests for minor edits, e.g., "Change the color of ... ", tend to be ignored. The AI returns the unmodified image.
The AI also heavily restricts the kinds of images it can generate or modify. It won't work on depictions of historical figures and stays conservative about adult content.
To be consistent with the images I have shown you before, I used a variation of my standard prompt: Generate an image of a group of physicists writing and discussing equations on a whiteboard. The equations should relate to quantum field theory. The image should be of photographic quality. The group of physicists should include both genders and show diversity.
In the past, you may recall that the equations were gibberish, the people tended to have messed-up hands, and there were weird things like using felt-tip markers on a chalkboard. Arms would have no hands.
In the image I have attached below, the generation is much better. There are many flaws, but it shows a plausible understanding of the prompt. The equations are still gibberish, but now the symbols and form are mostly correct, just not intelligible. There are spelling errors. The girl in the center has a broken finger on her right hand. The girl on the top left has a marker with a red end cap but a green tip. The guy on the right has a blue marker that writes in black.
On the whole, a credible job. Do you see any other errors or flaws?
I believe Adobe made a wise decision by partnering with Google to integrate Flash 2.5 into Photoshop and Firefly.
-- Robert Brown
Charles Gousha
Oct 17, 2025, 4:36:34 PMOct 17
to sv...@googlegroups.com
The man on the left is holding two pens. (Odd)
… and his right index finger ends at the first knuckle
I like the misspellings in the center.
Charles Gousha “We are the universe
Retired from Apple Inc . made manifest, trying to
user group lead – Silicon Valley MUG figure itself out."
RPG player, anime fan, SF nut – Delenn, Babylon 5
Retired from Apple Inc . made manifest, trying to
user group lead – Silicon Valley MUG figure itself out."
RPG player, anime fan, SF nut – Delenn, Babylon 5
--
You received this message because you are subscribed to the Google Groups "SVMUG" group.
To unsubscribe from this group and stop receiving emails from it, send an email to svmug+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/svmug/C1281ED0-B5FA-4331-8B5C-85199DE83681%40rlbrown.org.
<Gemini_Generated_Image_1wu53h1wu53h1wu5.png>
Reply all
Reply to author
Forward
0 new messages