Which model should I use?

Here is a list of models, what they make, what you might use them for, and any specialties they particularly excel at.

Generally, you get what you pay for: Higher-priced models tend to deliver better results and offer more capabilities.

See the individual model price list here: https://lumalabs.ai/pricing

Image Models

Uni-1.1 (Luma’s own)
A unified world model with extensive knowledge of the world and excellent grain structure. Does not support 4K.

Nano Banana family
Good for photographic, live-action looks, and excellent for animation and anime. Particularly strong for image editing, including changing, removing, or adding elements to an existing image. It may not always preserve the initial image verbatim during an edit.

GPT family of image models
Highly coherent, meaning objects generally look and behave as they do in the real world, such as producing correct five-fingered hands. Also highly adherent, meaning the models tend to follow instructions closely. Excellent for changing or editing elements within images. Has a contrasty grain structure that can build up and sometimes make an image look dirty or illustrative. People can also look haggard, although this can be caused by poor prompting habits.

Seedream family of image models
Highly coherent and adherent, with excellent grain structure. Produces very attractive images and has excellent image-editing capabilities.

Ideogram 4.5
Produces very clean images. Excellent for editing or changing specific elements in an image while leaving the rest of the image pixel-perfect. Also very good for non-photographic subjects.

Video Models

Ray family of video models (Luma’s own)
Strong, cost-effective general-purpose video models and excellent for 3D animation. They are a little old compared with some of the newer models below, so they are not quite as adherent or coherent or clean. Excellent for modifying existing video. Ray 3.2 can use up to 64 keyframes in both Create Video and Modify Video, and includes an HDR option. Excellent motion transfers.

Veo 3.1
The oldest premium model on Luma. Very good for animated subjects and excellent for animation and anime-style. Also very good for animated FX. The image can be somewhat noisy and less sharp than the other premium models. Very expensive for what you are getting.

Gemini Omni Flash
A low-cost model that is good at editing or changing elements within existing video. See the note at the end regarding Flash models.

Kling family of models
Kling 3.0 in particular is excellent for live-action footage, animation, and anime. A strong all-rounder that can also edit or change elements within an existing video. Excellent motion transfers. Very good at character consistency.

Seedance family of video models
Seedance is currently the king of the video models, and Seedance 2.5 is also the most expensive. Seedance 2.0 is almost as capable at roughly half the cost.

Seedance 2.5 is the only model on the market capable of generating a full 30-second video in a single render. It performs well across a wide range of subjects and particularly excels at live action. It is excellent at dialogue and character consistency.

2.5 can use up to 64 keyframes in Create. The Reference version can use up to 30 image references, 10 video references, and 10 audio references. Edit mode is excellent for video-to-video alterations and motion transfers. Some models include an HDR option that will appear if the stars align.

MiniMax family of models
H3 is the current flagship of the MiniMax group and falls just behind Seedance 2.5 in overall functionality for roughly one-third of the cost. It excels at live action and handles FX work better than most other models at any price. It follows motion paths better than the other models and is excellent at character consistency. Good video-to-video editing and motion transfers.

Flux 3
A low-cost, high-performing video model. Not quite on par with H3, but a strong option for the price.

Audio Models

ElevenLabs
Good for audio, dialogue, voice mapping, sound effects, and music.

Seed Audio
Excellent for dialogue and soundtracks. Speed and pitch control.

A Note on Flash, Mini, Fast, and Turbo Models

Models with these additions to their names often use a reduced instruction set and may have fewer steps in their completion process. This can manifest as reduced coherence and adherence: world physics may be less accurate, and it can be harder to get the model to follow your prompt or handle complex shots.

Because of this, these models are usually cheaper and faster to run. They can be excellent for idle animations or shots where you do not want the camera moving much or the character doing anything particularly complex.

However, while they are cheaper per generation, you may need to prompt many more times to achieve the result you want. In some cases, this can mean spending more credits overall than if you had used a more expensive, more adherent model and needed fewer attempts.