What changed through September 20
September adds more ways to develop an image after the first generation: compare several candidates, send them down different workflow branches, increase output resolution or extend the composition. This article covers September 1–20, 2026; it is not a full-month report.
For the earlier releases, read the July updates and August updates.
Generate several images and give each its own branch
Image nodes and the Create page now offer a Number of images setting from 1 to 4. Candidates appear as thumbnails beneath the result and in the enlarged preview. Click a thumbnail to choose the default output without running the task again.
Each candidate also has a numbered output port on the workflow node. Connect different candidates to different image or video nodes to explore several directions from one prompt. Connections made before generation follow the selected default output.
Generation is billed per successful image. If only some candidates succeed, those images are kept and only they are charged. Check the estimate for the selected quantity before running a batch.
Upscale a reference image or extend its composition
The Tools section of Add Node now includes Upscale and Outpaint. Choose an entry to create an image node with its mode already selected, then connect a reference image. Upscale uses that image as the basis for regeneration at a 2K or 4K output tier.
Outpaint lets you choose a target aspect ratio, expansion amount and direction. Use it to extend a composition for a horizontal banner or a vertical cover. Review the result for changes to the subject and fine details before using it.
These modes are available on supported sale-channel Nano Banana Pro and GPT Image models. The model selector indicates compatibility. Both modes generate one image at a time and are billed by the output tier; check current pricing for the selected model and channel.
More image models to choose from
Grok Imagine Image 2.0 joins the image model list with text-to-image generation, reference-image editing and combinations of up to five reference images. It offers 1K and 2K resolution tiers and low or medium quality settings.
GPT Image 2.5 Flare and GPT Image 2.5 Sunburst are separate choices for image generation and reference-image editing, with 1K, 2K and 4K tiers. These are provider resolution tiers: actual pixel dimensions can vary, particularly for Sunburst. Inspect the downloaded result if your delivery requires exact dimensions.
Bring files onto the canvas with a drop
Drag image or video files directly onto the workflow canvas. Files upload and become input nodes where you drop them; dropping several files places the nodes with space between them. You can then connect the assets to generation nodes.
Asset tag lookup and history filtering have also been improved. Unavailable nodes show more useful explanations, and fixes address missing image results on some 2K/4K channels and reference-image loading in local-storage deployments.
Clearer subscriptions, top-ups and introductory credits
The account balance page separates subscription and pay-as-you-go top-up options, with explanations of their use cases and improved mobile and keyboard navigation. Monthly plan credits apply to the current billing period; purchased top-up credits do not expire.
New accounts can receive introductory credits when the site has an eligible promotion enabled. The registration flow shows the configured amount, and billing records distinguish introductory credits from their expiry. Availability, amount and validity follow the offer shown at registration.
Private edition: larger uploads and finer model controls
Private-edition image uploads now accept files up to 50MB. Images above 10MB are compressed automatically, with the longest edge capped at 4096px, orientation corrected and transparency preserved where present. The Create page also accepts reference images pasted with Ctrl+V or Cmd+V.
Private administrators can enable or disable models locally, see menus according to their permissions and preview video generation records. These changes apply to private deployments; their availability depends on the installed release and administrator settings.
Find a guide, compare models or follow the next update
The website now includes a developer quickstart, a model outfit-change template guide and model comparison pages. Blog category filtering works again, and English, Simplified Chinese and Traditional Chinese RSS feeds let you follow new articles in a feed reader.
To try the image changes, start with one reference image, generate two candidates, select a preferred result and connect it to a supported Upscale or Outpaint node. Review that small run before expanding the workflow.