Forced sequential creation of images on cards
A
Aleksei Ilin
When creating presentations and web pages, I use the Nano Banana 2 mini image generation model.
When a lot of cards are created, the system first starts creating images one at a time (because few more cards have been created), but as a large number of cards are created, the system begins to generate images in parallel (i.e. 5-10 or more at the same time). This greatly reduces the quality. First of all, there are errors in words (and sometimes just an incoherent set of letters). My question is: is it possible to enable sequential image generation so that the AI starts generating the next image when the previous one is ready. Then the quality will be higher. If there is no such setting, I would be grateful if you would convey your wishes to the developers. In my opinion, this way the quality of the generation will increase and I won't have to spend credits later to regenerate the image several times.
I make up detailed descriptions, but due to the fact that there are a lot of pictures on the page, almost all of them have misspelled words. You have to spend a lot of time to create new ones. Moreover, the pictures themselves are good, but instead of words there is only a random set of letters.
Nik Payne (Gamma design)
Hey Aleksei Ilin, thanks for the super detailed writeup. That’s really helpful, and yeah I can see how parallel image gen would be brutal if it’s tanking text quality and burning credits.
I’ll pass this along to the team as a request for a “sequential image generation” option (only start the next image after the previous finishes). Quick couple questions so we can make it actionable:
1) About how many cards/images does it take before you notice it switching into parallel mode (roughly)?
2) When the misspelled words happen, are those words coming from your prompt (like you’re asking for specific text in the image), or is it text the model is inventing on its own?
If you have a screenshot of a page where it happens a lot, that would help too, but no worries if it’s a hassle.
A
Aleksei Ilin
Nik Payne (Gamma design)
1) It generates the first 2-3 images normally (i.e., 3 images is the limit), and then while it's creating these 3 images, it manages to create 10 more cards with text and starts creating images on all of them at once, and almost all of them have errors. And re-regeneration doesn't help, as it only makes things worse. It creates the best options the first time (if it hadn't made mistakes with the text). Then, I have to create separate images in Gemini and upload them manually.
2) For each image, there is a specific text that is written in the task with instructions to save it, and I also write what needs to be done in the instructions. This definitely works, but only on the first 2-3 images. I've attached an example. Almost all of the text is "abracadabra."
3) I also have a suggestion for creating posts for social media, where we create, for example, 10-15 cards. If it's a carousel post on Instagram, for example, it won't work because it uses the "editable text" + "image next to or on the background" model. As a result, the text is separated from the image, like two separate elements. For carousel posts, it should be a single image, for example, 1:1 in size. You have this functionality, but only when creating cards within the current general constructor. What I need is for it to create only cards that are just images (yes, you won't be able to edit the text, but you can simply regenerate the card). Nano Banana 2 and Nano Banana 2 mini handle these cards well. As a result, you'll have images ready to be uploaded.