In short
- For a persona, the mode matters far more than the model. Use image-to-video or reference-to-video.
- A 5-second image-to-video clip at default settings listed at $0.20 to $1.80 on 5 October 2026. Promotions that day cut the cheapest to $0.10.
- MiniMax H3 takes a LoRA in video, but only one trained for H3. Your image LoRA does not carry over, and WaveSpeed has no H3 trainer.
- Motion control copies the movement from a video of a real person. It costs more per clip, and it keeps the original sound unless you switch that off.
- No vendor publishes a number for character consistency, so you cannot rank models on the thing that matters most here.
On this page
- The modes matter more than the models
- What AI video costs per clip
- Using your character LoRA in video
- Motion control: the movement comes from a real person
- Higgsfield or WaveSpeed: subscription or pay per use
- Why there is no ranking of video models here
- How to test a video model on your character
- Questions people ask
A persona that only posts still images looks old-fashioned now. In six months it will look worse. That is why you need video.
Most guides to AI video do not help you with this. They rank models by how good a clip looks when you only type a text prompt. You already have a character, and you worked hard to get her face right. For you, one question matters: does she still look like herself after she moves?
Choose the mode first and the model second.
The modes matter more than the models
You can use each of these models in several modes. Some modes help you much more than others:
- Text-to-video makes up everyone in the clip. For a persona this is almost useless. You get a woman, but she is not your character.
- Image-to-video turns a first frame that you upload into a clip. You upload a picture of your character, so the face is right at the start. The model only has to move it.
- Reference-to-video takes a prompt and one or more reference images. Your character can appear in any scene you describe. She does not have to continue from one single frame.
- Video-extend adds more seconds to a clip you already have. You do not have to generate the clip again.
- Motion control (also sold as motion transfer or animate) takes your character image and a video of a real person moving. It moves your character the way that person moved.
You may have spent weeks on a reference sheetReference sheet A reference sheet is a fixed set of images of one character, from several angles and in neutral light. You use it as the reference for everything you generate afterwards. Open the glossary and a LoRA (a small add-on model trained on pictures of your character). The guide to character consistency is about that work. Text-to-video throws all of it away at the first step. Image-to-video, reference-to-video and motion control use it.
What AI video costs per clip
The pricing tables on the model pages bill these models per second of video, by resolution. A 5-second image-to-video clip at each page’s default resolution lists at: MiniMax H3 open weights (480p) $0.20, Wan 3.0 (720p) $0.50, Wan 3.0 Prime (720p) $0.75, Seedance 2.5 (720p) $1.80. Reference-to-video: MiniMax H3 $0.25, Wan 3.0 $0.50. MiniMax H3 image-to-video with a custom LoRA lists at $0.25. On 5 October 2026 the run button showed 50% off the MiniMax H3 open-weights models, 10% off Seedance 2.5 and 5% off Wan 3.0, so the four image-to-video clips cost $0.10, $0.475, $0.7125 and $1.62. Image edits list at $0.045 for Seedream V5.0 Pro Edit and $0.07 for GPT Image 2 Edit.
Wan 3.0 Text to Video advertises flexible durations from 2 to 30 seconds. Wan 3.0 Image to Video is described as animating a first-frame image into a cinematic video, and Wan 3.0 Reference to Video as creating coherent videos from prompts and multimodal references. ByteDance’s Seedance 2.5 is offered in image-to-video, text-to-video, video-edit and video-extend modes.
| Mode, list price | Cheapest listed | Mid | Premium |
|---|---|---|---|
| Image-to-video, 5 s | MiniMax H3, $0.20 | Wan 3.0, $0.50 | Seedance 2.5, $1.80 |
| Reference-to-video, 5 s | MiniMax H3, $0.25 | Wan 3.0, $0.50 | — |
| With an H3 LoRALoRA A LoRA is a small file trained on images of one subject. It teaches a general image model to reproduce that specific face or style. Open the glossary, 5 s | MiniMax H3, $0.25 | — | — |
| Motion control, per 5 s | Wan 2.2 Animate, $0.20 | Kling 3.0 Std, $0.63 | — |
| Image edit | Seedream V5.0 Pro, $0.045 | GPT Image 2, $0.07 | — |
To make the still you animate, ChatGPT AI influencer photos covers OpenAI’s image model with reference images and prompts.
The table shows list prices. Promotions on the run button come and go, so plan with the list price. WaveSpeed’s pricing page shows lower numbers again:
WaveSpeed’s pricing page lists video models by starting price per second: Seedance 2.5 $0.18, Seedance 2.0 $0.12, Veo 3.1 Fast $0.10, Kling 3.0 Std $0.084, Wan 3.0 $0.05, MiniMax H3 $0.04 and InfiniteTalk $0.03. It says starting prices are list prices at the lowest resolution or quality, and that the actual cost depends on resolution, duration and other settings.
Both are correct. The pricing page shows the cheapest setting, and the table shows each model’s default. Seedance 2.5 defaults to 720p, at twice the 480p rate ($0.36 a second against $0.18). Veo 3.1 Fast has sound on by default, which costs $0.15 a second on WaveSpeed.
Google’s Gemini API lists Veo 3.1 Fast video with audio, the default, at $0.10 per second for 720p, $0.12 for 1080p and $0.30 for 4k. The same page says the Veo 3.1 preview models, Fast included, are “deprecated and will be shut down on October 22, 2026”, and points to Gemini Omni Flash or “the GA models available through the Gemini Enterprise Agent Platform”.
Google AI for Developers — Gemini API pricing (opens in a new tab)checked
After 22 October 2026, Google’s own route to Veo 3.1 Fast is its Gemini Enterprise Agent Platform on Google Cloud. It lists the same rates as the Gemini API did.
“Veo 3.1 Fast”, “Video + Audio generation”: 720p “$0.10 / 1 count”, 1080p “$0.12 / 1 count”, 4k “$0.30 / 1 count”; “Video generation” without sound: 720p “$0.08 / 1 count”. The table does not name the unit; the Gemini API listed the same figures per second.
Google Cloud — Agent Platform pricing (opens in a new tab)checked
So Veo 3.1 Fast with sound at 720p costs a third less from Google directly than on WaveSpeed. You need a Google Cloud project with billing for it. The full list, with image models and dates, is on what generation costs.
The most expensive image-to-video clip in the table costs nine times as much as the cheapest. That matters because you will throw away most of your clips. At $0.20 you can generate forty clips, throw away thirty-six and still have spent only eight dollars. At $1.80 the same forty clips cost seventy-two dollars. A cheap model that you can run many times is better than an expensive one that you have to use sparingly. Before you pay for Seedance, check that it accepts your persona’s face at all: Seedance consistent character explains why it blocks realistic faces and the three official ways in.
Which model gives you the most usable clips right now? Ask it in #technical-inquiries. Someone there has usually had the same problem.
Using your character LoRA in video
One line in that price list looks more useful than it is. MiniMax H3 has image-to-video and reference-to-video versions that take a custom LoRA.
MiniMax H3 Image-to-Video LoRA accepts up to 3 LoRA weights per request, each given as a file URL and a scale.
WaveSpeedAI — MiniMax H3 Image-to-Video LoRA (opens in a new tab)checked
WaveSpeed’s guide to using a LoRA says the LoRA must exactly match the base model: the same model family, the same version and the same parameter size.
WaveSpeedAI blog — Use LoRA on WaveSpeedAI (opens in a new tab)checked
WaveSpeed’s training tools on 5 October 2026 included trainers for Flux Dev, Z-Image, Qwen Image, Wan 2.1, Wan 2.2 (images and image-to-video) and LTX-2 video. None was for MiniMax H3.
WaveSpeedAI — Training Tools collection (opens in a new tab)checked
So your character LoRA does not carry over. A LoRA you trained on FluxFlux Flux is a family of image-generation models from Black Forest Labs. It is widely used for photorealistic characters because it holds faces better than earlier open models. Open the glossary or Z-Image for your stills does not work in H3, and WaveSpeed cannot train an H3 LoRA for you today.
The closer route on WaveSpeed is Wan 2.2: a trainer for image-to-video LoRAs, and image-to-video models that load one. That is a second LoRA, made for video, on top of the one for your stills.
The Wan 2.2 I2V LoRA Trainer starts at $5 per run.
The price list only shows that these modes exist. It does not show how well they work. You have to test that yourself (see below). MiniMax H3 for AI persona video shows how to give it up to nine reference images of your persona, what a clip costs on MiniMax’s own API, and why the open weights are not licensed in the EU.
Motion control: the movement comes from a real person
In image-to-video the model has to make up the movement. Made-up movement is usually what shows that a clip is AI: a walk that floats, or hands that move in a way real hands cannot. Motion control skips that step. A real person does the movement on camera, and the model moves your character the same way.
Kling 3.0 Std Motion Control takes a character image and a driving video and generates a video in which the character performs the movements from the driving video. It costs $0.63 per 5 seconds, with a 3-second minimum charge and a maximum of 30 seconds, or 10 seconds if the character keeps the image’s orientation. The option to keep the driving video’s original sound is on by default. Wan 2.2 Animate does the same job at $0.20 per 5 seconds at 480p and $0.40 at 720p.
In the full plan comparison on its pricing page (under “See full comparison”, then “View More”), Higgsfield lists Kling 3.0 Motion Control at about 7 credits per 5 seconds at 720p and about 12 at 1080p, against about 7 and 8 for ordinary Kling 3.0 video.
For your persona this means three things.
The face still comes from your image. Motion control does not replace the work on a consistent face. It depends on it. If your reference frame is weak, you get a weak face that moves very realistically, and that looks worse than before.
It costs more per clip. Five seconds of Kling motion control costs $0.63. A 5-second MiniMax H3 image-to-video clip lists at $0.20. Pay the higher price when the clip is about the movement: a dance, a gesture or a walk towards the camera. Do not pay it for a slow head turn. Any image-to-video model can do that.
The driving video shows a real person, and it has their sound. The driving video is the video the movement is copied from. Record the movement yourself, or use footage you have the rights to. If you take a clip from someone else’s account, their performance ends up in your post. With the default setting, their sound does too. Switch the original sound off unless you recorded it yourself. Kling motion control for an AI persona shows how to bind her face, what a clip costs on each platform and how to film the driving video.
Higgsfield or WaveSpeed: subscription or pay per use
You can run these models on Higgsfield and on WaveSpeed. These are the two platforms mentioned most here, and each has its own review. They charge in very different ways. Which one is cheaper for you depends on how much you generate. The plan also decides whether you may sell the clips: AI video commercial use has the terms for Kling, Seedance and Hailuo.
In the full plan comparison, Higgsfield’s Starter plan is €19 per month billed annually and carries a fixed 270 credits per month, which it states is 135 Nano Banana Pro generations or roughly 15 Seedance 2.0 Fast videos. The Plus plan is €59 monthly, or €47 billed annually, for 1,200 credits, stated as 600 Nano Banana Pro generations or about 80 Seedance 2.5 videos. Starter includes Seedance 2.0 Fast and 2.0 Mini; Plus has access to all Seedance models. The plan cards at the top of the page give Starter as 135 Nano Banana Pro generations too, billed monthly or annually.
WaveSpeed uses a pay-per-use model with no monthly fees or commitments, and eligible new accounts receive $1 in free credits with no credit card required. Starting prices per image include Nano Banana Pro at $0.14, Nano Banana 2 at $0.045 and Z-Image Turbo at $0.005.
List price per image on the model pages: Nano Banana 2 at 1k resolution $0.07, Seedream 4.5 $0.04, Flux 2 Klein 4B $0.008. On 5 October 2026 both Nano Banana models were 10% off.
If you do the maths, you learn something that neither vendor’s marketing tells you. Higgsfield’s Starter plan is €19 for 135 Nano Banana Pro generations. That is about €0.14 each. WaveSpeed lists the same model at $0.14, and you only pay for what you generate. One price is in euros and the other in dollars, so the two are not exactly the same. They are close enough to show this:
The subscription does not make each generation cheaper. You pay for the interface, the presets, and for never having to use an API.
That is a fair thing to pay for. For most people who are starting out it is the right choice. You should still know what your money buys. Some people generate in bursts: nothing for three weeks, then two hundred frames in one weekend. For them a subscription is a bad fit, because the monthly credits they do not use are lost. Paying per generation costs them less for the same output. If you generate a steady amount and do not want to think about billing, the subscription is fine. If you pay per generation and render in batches, Seedance with Claude Code or Codex shows a script that quotes every clip and stops at a spending cap.
The two platforms differ most on the cheap models. On WaveSpeed, Z-Image-Turbo costs half a cent per image and Flux 2 Klein costs under one cent. At those prices you can afford to generate a lot of images and keep only the best ones. A credit system that is priced around premium models does not give you those low prices in the same way.
Why there is no ranking of video models here
What the sources do not say
Ranking these models on published numbers
This article was going to end with a ranking. It does not, because the numbers that vendors publish are not enough to build one.
Three things stopped it. First, the prices use different units. The same catalogue lists some video models per second and others per generation. You cannot compare two prices until you have checked which unit each one uses. Second, the descriptions of what a model can do are marketing. They are not measurements. “Hollywood-grade cinematic” is in one product description, and you cannot test a model against that. Third, and most important, no vendor publishes anything about character consistency at all. There is no metric, no benchmark and no definition. Character consistency decides whether you can use a model for a persona, and it is missing from every spec sheet.
A ranking would only have been my impression of a few clips. What is left is the dated price table above and one piece of advice: test on your own character. None of these numbers can tell you how a model handles her.
How to test a video model on your character
- Take one reference frame of your character. Pick the strongest one you have: front-on, in neutral light.
- Run the same frame through two models in image-to-video. Use the cheapest model and a mid-priced one, with the same prompt and the same length. At list prices of $0.20 and $0.50 for 5 seconds, that costs under a dollar.
- Watch only the face, and only at the start and the end of the clip. It is easy to judge how good the movement looks, but that matters little here. What matters is driftDrift Drift is the slow change in a character’s face over time, until images made weeks apart no longer look like the same person. Open the glossary: how much the face changes between the first and the last frame. Drift decides whether you can post the clip.
- Do it again with a harder frame: a three-quarter angle, or hair across the face. Every model handles the easy frame, so the easy frame tells you nothing.
- Keep the seedSeed A seed is the number that sets the random starting point of a generation. The same prompt with the same seed produces the same image. Open the glossary and the prompt. In two months you can run the same test on a newer model and compare the results fairly.
That is four generations and about ten minutes. It tells you more about your character than any ranking article can.
If the face still changes between still images, fix that before you spend money on video. The guide for that is character consistency. The glossary explains the terms used above: LoRA, seed, reference sheet and drift. Video also adds hours and cost to your week. To plan for that, read where the hours go and what an AI persona costs to start.
Sources for this article
The platform rules, laws and numbers above come from these documents. The date says when we last opened each one and checked it.
- WaveSpeedAI model cataloguechecked
- WaveSpeedAI pricingchecked
- Google AI for Developers — Gemini API pricingchecked
- Google Cloud — Agent Platform pricingchecked
- WaveSpeedAI — MiniMax H3 Image-to-Video LoRAchecked
- WaveSpeedAI blog — Use LoRA on WaveSpeedAIchecked
- WaveSpeedAI — Training Tools collectionchecked
- Higgsfield pricingchecked
Questions people ask
Which AI video model is best for an AI persona?
The mode matters more than the model. Use image-to-video or reference-to-video. Both start from a picture of your character, so the model does not have to invent the face. A cheap model in one of these modes beats an expensive model in text-to-video.
How much does one AI video clip cost?
On WaveSpeed on 5 October 2026, a 5-second image-to-video clip at default settings listed at $0.20 on MiniMax H3 (480p), $0.50 on Wan 3.0 (720p) and $1.80 on Seedance 2.5 (720p). Promotions that day cut them to $0.10, $0.475 and $1.62. Prices in this field change, so look at the date next to every price, including this one.
Can I use my character LoRA in video?
Only a LoRA trained for that video model. MiniMax H3 on WaveSpeed takes up to three LoRA files in image-to-video and reference-to-video. But WaveSpeed says a LoRA must match the base model exactly, and on 5 October 2026 it had no trainer for H3. A LoRA you trained for Flux or another image model does not work there. WaveSpeed does have a trainer for Wan 2.2 image-to-video LoRAs, at $5 a run.
How long can a generated clip be?
Wan 3.0 advertises clips from 2 to 30 seconds in a single generation. Several models also have a video-extend mode. It adds more seconds to a clip you already have, so you do not have to generate the clip again.
What is motion control in AI video?
It is a mode that takes two inputs: a picture of your character and a video of a real person moving. The model moves your character the way the person in the video moved. On WaveSpeed on 5 October 2026, Kling 3.0 Std Motion Control cost $0.63 per 5 seconds, for clips of up to 30 seconds.
Does adding video mean my persona needs a voice?
No, it does not have to. Most social video in this niche is watched muted or with music over it. Lip-sync and avatar models are a separate group of tools. They solve a different problem from turning a still image into a clip.