Google’s January 2026 update did not add a model. It rebuilt one capability — Ingredients to Video, the one that turns reference images into a clip — and pushed it onto surfaces it had never reached.
One naming caution: the same three words label an unrelated feature from another vendor.
What the update actually changed
The definition is unchanged: Ingredients to Video is Google’s “capability that lets you create videos based on reference images.” What changed is the output.
Effort first. The update “makes videos more expressive and creative, even with simple prompts,” and “even with short prompts, you can generate dynamic and engaging videos based on ingredient images.”
Then consistency, in two claims. Identity: “Keep your characters looking the same even as the setting changes, making it easier to tell a full narrative by having the same character appear across multiple scenes.” Blending: “Combine disparate elements — like characters, objects, textures and stylized backgrounds — into a cohesive, high-impact clip.”
One thing is missing, and it is what guides keep inventing. No Google page publishes a maximum number of ingredient images; third-party write-ups state a ceiling of three as if it were specification. Read the live limit off the Flow panel. The prompt guide covers how those images get written into the shot.
Three image-driven paths in Flow
Flow starts from a picture in three ways, and picking the wrong one is the usual reason an image-driven generation disappoints.
One reference image
The narrowest path. Google’s model page describes giving Veo “reference images of a scene, a character, or an object to guide its generation.” Use it when exactly one thing must survive.
Frames to Video
A first frame, a last frame, and the movement generated between them. That is interpolation, not composition, which is why it is the wrong tool for “this person in that place” — both images are one scene at two moments. The getting started guide covers the same toolset.
Ingredients to Video
Combinatorial instead. You supply a character, an object, a texture, a stylised background, and the prompt says how they meet. The January update added a consistency promise to each slot, which is what makes a multi-scene sequence possible.
The rule: two moments of one scene means Frames; two things that share one scene means Ingredients.
Prepare the ingredient set first
Most disappointing generations are decided before the prompt box is touched.
Isolate each element. A portrait with a busy street behind it gives the model two competing answers to “what is the subject here.”
Make the lighting agree. A subject shot in flat studio light, dropped into a night scene, reads as pasted on. Choose references that already agree, or name the lighting so the model can relight.
Write the mapping down before uploading:
image 1 — character: who it is, cropped tight, one lighting direction
image 2 — object: what the character holds or uses
image 3 — setting: where, plus the texture or grade to reuse
prompt sentence: <image 1> + <image 2>, in <image 3>, then the actionGoogle’s own advice is to generate those slots. The announcement’s pro tip is to “use the new Nano Banana Pro (Gemini 3 Pro Image) in the Gemini app or Flow to create your ingredient images.” A character sheet built with an image model first is a documented habit, not a workaround.
Two worked examples
Both prompts describe eight-second shots, the only length Google publishes: “Veo videos are 8 seconds long.”
Example one: one street scene
Ingredients
image 1 — a woman in her thirties, cropped at the shoulders, soft window light from the left
image 2 — a rain-slick city street at night, neon reflections in standing water
image 3 — a folded paper map, warm paper texture, no legible text
Prompt
The woman from image 1 walks along the street from image 2, holding the map from image 3.
Medium shot, slow tracking with her, shallow depth of field. She stops, looks up from the
map, and says "This is the street." Audio: night ambience, footsteps in shallow water.Example two: vertical product beat
Ingredients
image 1 — a ceramic teapot, shot on white, hard key from the right
image 2 — a linen tablecloth texture, natural daylight
image 3 — a pale kitchen wall with a window, morning light from frame left
Prompt
Close-up of the teapot from image 1 on the linen from image 2, against the wall from image 3.
Vertical 9:16 framing, static camera, slight handheld drift. Steam rises, the lid lifts once
and settles. Audio: a low boil, a lid ticking, quiet room tone.We have not run either prompt and claim no result. Each block follows the mapping above: name the ingredient, give it a job, spend the rest on the action no image can contain. Sound works as in any Veo clip, which the audio prompting guide sets out.
Sequences, not single clips
Consistency across settings only pays off past the first shot, and Extend is how you get there. The Flow announcement describes videos “lasting for a minute or more, that connect to and continue the action from your original clip,” each one generated from the final second of the previous clip. Hold the reference set constant and the character keeps its face while the setting changes.
The published limit still applies. Google states that “creating videos with natural and consistent spoken audio, particularly for shorter speech segments, remains an area of active development,” so keep spoken lines short if a sequence depends on them.
Vertical output and 4K upscaling
Google writes that “for the first time, ‘Ingredients to Video’ supports generating videos in a native 9:16 aspect ratio,” producing “full-screen vertical storytelling without cropping or quality loss” — a native render, not a cropped landscape frame.
Upscaling carries a footnote that decides where you can work: “upscaling to 1080p and 4K is only available in Flow, the Gemini API and Vertex AI.” The API and pricing guide covers the route where resolution is billed rather than clicked.
Availability splits the same way. The update is “launching in the Gemini app, YouTube, Flow, Google Vids, the Gemini API and Vertex AI,” and clips carry SynthID.
Frequently asked questions
How many reference images does it accept? No Google page publishes a number. Third-party guides say three and treat it as specification; the official wording is only “multiple reference images.”
Is it the same as Frames to Video? No. Frames bridges two moments of one scene. Ingredients assembles different elements into a scene that existed in none of the source images.
Can I get vertical or 4K out of it? Both are documented. Native 9:16 arrived with the January 2026 update, and 1080p and 4K upscaling is limited to Flow, the Gemini API and Vertex AI.
Do the references already have to look like one scene? Nothing published says so, but references whose lighting agrees need less work. The consistency claims describe holding an element steady, not reconciling two contradictory ones.
Pick the mode by counting moments rather than elements, give each image one job, and the prompt has almost nothing left to do except the action.