Ceilings run from fifty on Seedance 2.5 down to one on Hailuo 2.3 and Kling 3.0 Turbo, with Veo 3.1 Lite having no references section at all. But the count is the less useful half. What each model accepts differs too — Wan 2.6 won't take an image, Grok Imagine 1.5 won't take a clip, and four models are images only.
"How many references" is the question people ask. "What kind" is the one that stops a workflow, and it's answered in a line of grey text most people never read.
Why does the ceiling matter?
Because references are how you hold a character, a place and a look across shots.
A reference is what makes a generation a continuation rather than a fresh guess. Attach a saved actor and the model is being shown who this is; describe them instead and it's inventing them again. Same for a location, a style, a prop.
So a model's reference ceiling is a ceiling on how much of your world you can carry into a single shot. On a series with a recurring cast, a recurring room and a fixed look, that adds up faster than you'd expect.
A reference is the difference between showing the model and telling it. The ceiling is how much you can show at once.
What does every model accept?
All eighteen, with the ceiling and the accepted types.
| Model | Ceiling | What it accepts |
|---|---|---|
| Seedance 2.5 | 50 | Images, clips or audio |
| Seedance 2.0 | 15 | Images, clips or audio |
| Seedance 2 Fast | 15 | Images, clips or audio |
| Seedance 2 Mini | 15 | Images, clips or audio |
| MiniMax H3 | 9 | Images, clips or audio |
| HappyHorse 1.1 | 9 | **Images only** |
| Kling 3.0 Omni | 7 | Images or clips |
| Vidu Q3-Turbo | 7 | **Images only** |
| Wan 2.7 | 8 | Images, clips or audio |
| Gemini Omni Flash | 6 | Images or clips |
| Wan 2.6 | 3 | **Clips or audio — no images** |
| Veo 3.1 | 3 | Images or clips |
| Veo 3.1 Fast | 3 | Images or clips |
| Grok Imagine 1.5 | 3 | **Images or audio — no clips** |
| LTX-2.3 Fast | 2 | Not shown in the panel |
| Hailuo 2.3 | 1 | **A single image** |
| Kling 3.0 Turbo | 1 | **A single image** |
| Veo 3.1 Lite | **None** | No references section at all |
Source: Hexcoded Creative Studio picker, Video → Create, September 2026. Accepted types taken from each model's own reference hint text.
Five things worth pulling out.
Seedance 2.5 is an outlier by a wide margin. Fifty references against fifteen for the next tier, which is more than three times the nearest competitor on the roster.
Wan 2.6 won't take an image. Its hint reads "add clips or audio," and its description is "restyle a clip you already have." Those fit together — it's a restyling model, so the input is footage rather than stills.
Grok Imagine 1.5 won't take a clip. Images or audio only, which is unusual for a model whose description names synchronised audio.
Four models are effectively image-only. HappyHorse 1.1 and Vidu Q3-Turbo say "add images." Hailuo 2.3 and Kling 3.0 Turbo say "add an image" — singular, matching their ceiling of one.
Veo 3.1 Lite has no references section at all. Not a ceiling of zero; the field isn't in the panel. So there's no character consistency route on that model.
Where this falls short. LTX-2.3 Fast's ceiling is two, but its hint text wasn't visible in the panel we read, so the accepted types are unknown. Rather than infer them from the model's description, the table says so.
What does the fifty actually break down into?
Not fifty images, and this is the figure most often misquoted.
Hexcoded's own Creative Studio documentation gives the composition: up to 30 images, 30 actors, 10 videos and 10 audio, with the combined total capped at 50.
So "fifty reference images" is wrong twice over. It treats a total as a category, and it omits actors entirely — which is the category most likely to matter, since that's where your cast lives.
We've made that error ourselves in an earlier internal reference, which is why it's worth stating plainly.
Fifty is a total, not an image count. Thirty images maximum, and thirty of the fifty can be actors.
Does a higher ceiling mean better results?
No, and it's worth being direct about that.
The ceiling is what the model will accept. Whether it renders well with thirty references attached is a different question, and nothing in the picker addresses it.
Attaching fifty references to one generation is technically available and almost certainly a bad idea — you're giving the model fifty things to reconcile in a single shot.
The practical use of a high ceiling isn't volume. It's not having to choose. On a scene with a character, a location, a style and a prop, four references is plenty and a ceiling of three would force a compromise.
How does this interact with the shared pool?
One thing worth knowing, because it changes what the ceiling means in practice.
Creative Studio maintains the reference pool across Create, Edit, Extend and Motion within an active session rather than resetting after each generation. So references attached to a shot are still attached when you extend or edit it.
That's useful, and it means a model's ceiling applies to whatever is in the pool at the time rather than to what you deliberately added for this shot. If you've moved to a different scene in the same session, references from the previous one may still be counting against it.
Worth checking what's attached before you generate rather than after.
What about image generation?
A separate and much lower ceiling.
Creative Studio's image engine takes 16 references against video's 50. And in Image Edit mode, video and audio reference tabs are disabled entirely — the reference set is visual only.
So a reference set assembled for video work doesn't transfer wholesale to image work, and a reference-heavy image workflow has a third of the room.
How should you choose on this basis?
Three questions, and the first eliminates the most models.
What type of reference do you actually need?
If it's a clip, Wan 2.6 works and HappyHorse, Vidu, Hailuo and Kling Turbo don't. If it's audio, Grok and the Seedance family take it and Veo doesn't. Type eliminates more models than count does.
How many distinct things carry across the shot?
A character, a location, a style and a prop is four. Most models handle that. A ceiling only bites above about eight, which is where the Seedance family and MiniMax H3 separate from the rest.
Does the character need to persist?
Then avoid Veo 3.1 Lite entirely — no references section means no way to attach a saved actor. Veo 3.1 or Veo 3.1 Fast both take three, which is enough for a character plus a frame.
All ceilings and accepted types read from Hexcoded's Creative Studio picker in Video → Create mode in September 2026, with each model's reference field observed individually. Composition figures from Hexcoded's Creative Studio documentation. Limits change with model versions and differ by mode. Verify in the picker before relying on this.
- Ceilings run from 50 on Seedance 2.5 to one on Hailuo 2.3 and Kling 3.0 Turbo. Veo 3.1 Lite has no references at all
- The type matters more than the count. Wan 2.6 won't take an image; Grok Imagine 1.5 won't take a clip
- Four models are image-only: HappyHorse 1.1, Vidu Q3-Turbo, Hailuo 2.3 and Kling 3.0 Turbo
- Seedance 2.5's fifty is a total — 30 images maximum, 30 actors, 10 videos, 10 audio. Not fifty images
- A higher ceiling means not having to choose, not better output. Fifty references in one shot is a bad idea
- The pool persists across modes within a session, so it counts what's attached rather than what you added today
- Image editing takes 16 rather than 50, and disables video and audio references entirely
- If a character has to persist, don't use Veo 3.1 Lite. There's nowhere to attach it
From fifty on Seedance 2.5 down to one on Hailuo 2.3 and Kling 3.0 Turbo. The middle tier is fifteen on Seedance 2.0, 2 Fast and 2 Mini, then nine on MiniMax H3 and HappyHorse 1.1. Veo 3.1 Lite has no references section at all.
No. Fifty is the combined total, composed of up to 30 images, 30 actors, 10 videos and 10 audio. So "fifty reference images" treats a total as a category and omits actors, which is usually the category that matters most.
Wan 2.6, whose reference field says "add clips or audio." That fits its description — it's a restyling model, so the input is footage rather than stills. Every other model with a references section accepts images.
Four. HappyHorse 1.1 and Vidu Q3-Turbo say "add images," and Hailuo 2.3 and Kling 3.0 Turbo say "add an image" — singular, matching their ceiling of one. Grok Imagine 1.5 takes images or audio but not clips.
It means not having to choose rather than better output. A scene with a character, a location, a style and a prop needs four references, which most models handle. Attaching fifty to one generation is available and almost certainly counterproductive.
No, much lower. Creative Studio's image engine takes 16 references against video's 50, and Image Edit disables video and audio references entirely. So a reference set built for video work doesn't transfer wholesale to image work.
Ceilings visible before you generate
Every model shows its reference count and what it accepts before you attach anything. Eighteen video models on one credit balance, and the pool carries across Create, Edit, Extend and Motion.
Open Creative StudioMore on model capability, access and rights in Models.