Radamés Ajna — testing nano-banana with recursive editing as long video gen

TL;DR

Radamés Ajna demonstrates using Nano-Banana (Gemini 2.5 Flash Image / Nano-Banana 1) as an autoregressive video generator by recursively feeding edited frames back into the model with the prompt “Show this scene one second in the future.” The result is a chained image-editing loop that produces a frame sequence resembling a short video, without any explicit video model — turning a still-image editor into a poor-man’s frame-by-frame world model. The tweet went semi-viral (375K views) and is being shared as a curiosity in the research community.

Key claims

A single-image editor (Nano-Banana) can be coerced into producing video-like output by recursive self-conditioning with a “one second in the future” prompt [tweet body].
The prompts used are simple natural-language scene-progression instructions, with no external motion controls, depth, or video model in the loop [tweet body].

Method

The author iteratively edits an image with Nano-Banana, each time prompting “Show this scene one second in the future.” The output of step n becomes the input image of step n+1, with no other state passed between calls. Effectively, the model is asked to simulate one second of dynamics per step from a still frame, and the chain is repeated to extend the “video.” This is an instance of the classic autoregressive-rollout pattern applied to an image editor that wasn’t trained as a video model.

Results

No quantitative numbers — this is a demo tweet, not an experiment. The attached video (not retrievable as a frame stream from the text scrape) is reported by the author to show recognizable scene progression. The thread does not address drift, identity loss, or motion plausibility, which are the obvious failure modes one would expect from iterative image-editor rollouts without temporal training.

Why it’s interesting

This is a folk-recipe analog to the rigorously studied autoregressive video DiTs in the wiki: Rolling Forcing: Autoregressive Long Video Diffusion in Real Time and Self-Forcing++: Towards Minute-Scale High-Quality Video Generation both attack exactly this train-test gap where naive rollouts of an image/short-video model accumulate error and drift. The tweet also speaks to the broader bet captured by Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing and Nano-consistent-150K — Identity-Consistent Image Editing Dataset Built with Nano-Banana — Nano-Banana’s identity preservation across edits is good enough that people are now using it as a generative substrate well beyond single-step editing. The contrast with purpose-trained autoregressive video models is the interesting bit: those models exist specifically because chained image editing has unknown drift characteristics at length.