Whisk took three pictures instead of a sentence. One said what the image was of, one said where it was, one said what it looked like. It combined them. That is the whole idea, and it is worth understanding even though the tool itself was retired on 30 April 2026 — because the same mechanism now runs inside Flow.
What each input controlled
Subject set the main focus: a person, an object, an animal, a vehicle. Whisk pulled the defining characteristics out of this image rather than copying it, which is why a subject came back recognisable but rarely identical.
Scene set the environment — the place, the light, the background the subject was dropped into.
Style set the treatment: palette, medium, mood. In practice this input had the heaviest hand on the final look, and a style reference that mixed several aesthetics produced a muddle, because there was no single signal to copy.
The reason for splitting them is that in a written prompt all three are tangled together. Change the style in a sentence and the subject usually shifts with it. Keeping them in separate slots let you swap one and hold the other two still.
Why the inputs mattered more than the interface
The mechanism was simple enough that results came down almost entirely to what you fed it. Three references that already share compatible lighting and colour combine cleanly. Three that do not, will not — no amount of regenerating fixes a mismatch you introduced at the input stage.
The useful habit was reading the description Whisk generated back after each upload. That text showed what the model thought it was looking at. If the description was wrong, the output was going to be wrong, and the fix was a different reference image rather than another attempt.
Where the mechanism lives now
Google folded Whisk into Flow on 30 April 2026. The three-input concept survives there as Ingredients — reference images that hold a subject or character consistent — but it is aimed at video rather than single stills, appearing in the feature list as "Ingredients to Video".
The models changed with the move. Whisk generated images with Imagen 3, steered by Gemini. Flow generates images with Nano Banana (previously ImageFX), video with Veo 3.1, and handles reference-based edits with Gemini Omni. So an older tutorial naming Imagen 3 was accurate when it was written and is not accurate now.
Flow also adds things Whisk had no equivalent for: Scenebuilder, a timeline for extending a shot or joining clips with continuous motion, plus text-to-video, frames-to-video, video extension, image upscaling, Characters and Avatars.
What it costs
Whisk was free. Flow starts free at 50 credits a day, with paid tiers above that — Plus at $4.99/month, Pro at $19.99/month and Ultra higher still. Video generation consumes credits considerably faster than images, so a free allowance that felt generous under Whisk runs down quickly if you are producing clips.
What carries over
The buttons moved; the judgement did not. Pick clean, uncluttered references. Check what the model says it sees before you generate. Keep lighting and colour compatible across your inputs. Expect approximation rather than reproduction — and if you need something to look exactly like itself, this was never the right tool and still isn't.
Read the full migration breakdown
Feature names, models and pricing from the Flow product page and Google Workspace Updates. Checked 4 September 2026. Google changes these regularly — verify before relying on the figures.