← All articlesTools

AI keyframes: why the stills come first in AI video

Nano Banana Pro, GPT Image 2, Seedream 5.0, Midjourney V8: how stills are used as keyframes in AI video, from character consistency to resolution.

An AI keyframe is a still image, generated with a model such as Nano Banana Pro, GPT Image 2 or Seedream 5.0, that locks the composition, light and characters of a shot before it is animated. The video engine uses it as a first or last frame, or as a reference: as of November 2026 Kling 4.0 accepts up to 10 keyframes in a single clip, and Veo 3.1 a start frame and an end frame.

Starting from stills costs less, gets sign-off sooner and gives the video a foundation that a prompt alone cannot guarantee. This guide covers how engines read keyframes, which image models to use, how to keep characters consistent and what resolution to work at, with versions checked in November 2026.

What an AI keyframe is and how a video engine uses it

In AI video the word keyframe covers three different uses, and it pays to tell them apart before choosing an engine.

  • First frame. The image the clip starts from, in image-to-video mode. The engine animates it and, in theory, stays faithful to it.
  • First and last frame. Two images, and the engine works out the path between them. On Higgsfield, Veo 3.1 does this in clips of 4, 6 or 8 seconds at 24 frames per second.
  • In-between keyframes. Several images pinned to specific points on the timeline. Kling 4.0 takes up to 10 and lets you fix a pose, a product angle or a line of text at an exact second; Kling 3.0 stopped at start and end.

Then there are reference images, which are not frames at all: they tell the model who the character is or what the product looks like, not where it sits in the shot. Veo 3.1 accepts up to three, Kling 4.0 up to 15 across images and videos, and Cinema Studio 4.0 up to 50. Editing tools work the same way: in Firefly, the Edit Video feature of Luma Ray3.14 accepts first and last frames or reference images.

Why stills come first in AI video

An image regenerates in seconds. A clip takes longer and costs more credits: in August 2026, a 10-second 1080p clip with Seedance 2.5 on Higgsfield cost 120 credits. But cost is not the main reason; control is. Composition, lens, the colour of the light, product details and on-screen text are easier to decide on a still frame, where you can study them, fix them and get them approved.

It is also how clients understand a project. In the pre-production act of our method, the look, digital casting, locations and storyboard are signed off as images before a single frame is animated. If the client changes their mind about the colour of a jacket, that is a fix on a still, not a scene to regenerate.

Here is a documented example. In the breakdown of a one-minute short published by Higgsfield, the team first built the characters with GPT Image 2 at 4K on a grey background, the locations with Soul Cinema and the props with Nano Banana Pro, because GPT Image 2 rendered them flat. Only then did they use them as references for the video. According to the post, the location image did more than anything else to determine the quality of the clip.

Image models for AI keyframes in November 2026

Versions change fast. This table cross-checks Adobe’s partner model list, updated on 23 October 2026, against Higgsfield’s image page. The strengths listed are those claimed by the developers or the platforms.

ModelDeveloperClaimed strength
Nano Banana ProGoogleMultiple references, same subject from different angles, text
Nano Banana 2GoogleConsistent characters across images, targeted edits
GPT Image 2OpenAIEdits that preserve faces and composition
Seedream 5.0ByteDanceNative 2K, edit by example
FLUX.2Black Forest LabsMultiple references, typography
Midjourney V8MidjourneyCharacter and style references
Soul 2.0 and Soul IDHiggsfieldPhotographic look, trained face
Firefly Image Model 5AdobeNative 4MP, “commercially safe”

A few practical notes. In Firefly, Nano Banana Pro takes up to six reference images and merges them into one; FLUX.2 takes up to four. Adobe’s list already includes two GPT Image 2.5 variants, a sign that the next version is out there. For packaging, signage and logos, text rendering matters as much as faces.

Midjourney reached V8 in March 2026 and V8.2 in July, according to the timeline on Wikipedia. It does not appear on Adobe’s or Higgsfield’s lists, so it runs as a separate workflow. Firefly Image Model 5 generates at native 4 megapixels and is the only model in the table that its developer calls “commercially safe”.

Consistency: same face, same product, same light

A beautiful keyframe is not enough: it has to match the next one. Three habits keep things consistent.

  • Character sheets. For each character, a clean close-up and a full-length shot on a neutral background. These are the images Higgsfield’s breakdown found most effective.
  • Saved references. Platforms now treat them as named objects. In Firefly, Kling elements take up to four images labelled as character, object or location, and Adobe has Elements in private beta across Firefly. Higgsfield has Soul ID, which according to designboom works best with around twenty photos.
  • One model per sequence. Every model has its own skin, grain and light. Switching mid-scene shows, even when the face is right.

For the harder cases, such as a character who returns across ten episodes, see our guide to consistent characters in AI video.

What resolution to generate AI keyframes at

The rule: at least the video’s resolution, in the final aspect ratio. As of November 2026, many engines on Higgsfield top out at 1080p, from Veo 3.1 to Kling 4.0 and Cinema Studio 4.0. A 2K keyframe, which Higgsfield lists as native for Seedream 5.0 and Nano Banana Pro, is more than enough. GPT Image 2 reaches native 4K, according to the same platform.

  • Generate in the delivery format. A 16:9 keyframe cropped to 9:16 loses more than half its pixels and nearly always its composition. If you need both formats, make two keyframes.
  • No creative upscaling on keyframes. Upscalers that invent detail can alter a logo or a line of text. A clean 2K keyframe beats a 4K one with made-up detail.
  • 4K comes last. Delivery resolution is reached on the edited video, as we explain in our guide to AI video upscaling.

From keyframes to video: a six-step workflow

  1. Start from an approved storyboard: each panel becomes a keyframe. More in our guide to the AI storyboard.
  2. Build character sheets, locations and props, and save them as named references.
  3. Generate the keyframes for every shot with the same model, in the final aspect ratio.
  4. Get the keyframes signed off by the client before animating.
  5. Pick the engine for the kind of control you need: first and last frame, multiple keyframes or references only.
  6. Generate variants, select the best take, then move into the edit.

A keyframe is a directing decision made in advance. The more decisions you make on the still, the fewer you leave to the engine’s whims. It is the same principle behind our work on digital characters and series: first you establish who the character is, then you make them move.

Frequently asked questions

What is the difference between a keyframe and a reference image?

A keyframe is a frame of the clip: at the start, at the end or at a set point. A reference image never appears in the video; it tells the model what a character, product or style looks like.

Which image model is best for keyframes?

There is no single winner. Pick the model that renders your subject best, whether a face, a product or a line of text, and use it for the whole sequence: consistency matters more than any single image.

Can you make an AI video without keyframes?

Yes, with text-to-video the engine generates everything from the prompt. It is fine for exploring ideas, but for commissioned work it gives you less control over composition, characters and product details.

Sources

  1. Veo 3.1 on Higgsfield · Higgsfield, 2026-11
  2. Kling 4.0 intro · Higgsfield, 2026-11
  3. Meet Cinema Studio 4.0: New Generation of AI Filmmaking · Higgsfield, 2026-08-12
  4. Seedance 2.5 on Higgsfield · Higgsfield, 2026-08-06
  5. Recreate our 1-minute Seedance 4K film: the full breakdown with all prompts · Higgsfield, 2026
  6. AI Image Generator · Higgsfield, 2026-11
  7. GPT Image 2 · Higgsfield, 2026-11
  8. Seedream 5.0 · Higgsfield, 2026-11
  9. Partner models in Adobe products · Adobe Help Center, 2026-10-23
  10. Create with top AI models like Nano Banana Pro and Firefly Image Model 5 directly in Adobe apps · Adobe Blog, 2025-11-20
  11. Google Gemini (Nano Banana) in Adobe Firefly · Adobe, 2026-11
  12. GPT Image in Adobe Firefly · Adobe, 2026-11
  13. FLUX models in Adobe Firefly · Adobe, 2026-11
  14. Kling AI in Adobe Firefly · Adobe, 2026-11
  15. Luma AI in Adobe Firefly · Adobe, 2026-11
  16. Adobe Firefly introduces new agentic capabilities and an upgraded creative AI studio · Adobe Blog, 2026-06-18
  17. Adobe MAX 2025: Firefly Image Model 5, video editor, Generate Soundtrack and Speech (press release) · Adobe via Business Wire, 2025-10-28
  18. Midjourney · Wikipedia, 2026-11
  19. Higgsfield brings art-directed quality to AI image generation at production scale · designboom, 2026-04-03

Got an idea? Let’s make it a film.

We’ll reply with a concrete proposal, built around your brief.