Can I choose the model myself?

Yes. The agent proposes a model per job and says why; you can override it on any row before running. Autopilot, when you enable it, lets the agent decide and run by itself.

Do you add new models when they come out?

Yes, once they win their category in blind human votes and in our own production. A model can also be removed when it is discontinued by its maker or overtaken. The list on this page is the live one.

Do all models accept real faces and products as references?

Most do; some refuse real faces unless rights are verified. Miravela routes the job to a model that accepts your reference, or tells you why a model refused it, and you can use another one.

Models

Which AI image and video models does Miravela run?

12 hand-picked models — 5 for images (Nano Banana 2, GPT Image 2, MAI Image 2.5, Seedream, Flux Pro) and 7 for video (PixVerse C1, Veo 3.1, Kling, Grok Imagine 1.5, Wan 2.7, Seedance, Hailuo) — every one of them at its maximum quality, on every plan, with no per-model surcharge. The built-in agent picks the model for each job; you can always override it.

The image studio: the agent, the brand kit, and autopilot off by default

Why twelve, not fifty

Every model on this list is here because it is the best at one kind of job a brand actually orders: editorial realism, packshots, typography inside the image, cinematic stills, runway movement, brand films with sound, expressive faces. A catalogue of fifty models is a catalogue you have to know; a list of twelve, each with a job, is a list the agent can choose from for you.

The list changes as the field moves. When a new model wins its category in blind human votes and in our own production, it comes in; when one is discontinued or overtaken, it goes. The blog says what changed and why.

Maximum quality, by rule

Each model runs at its own highest resolution and setting by default — 4K where the model does 4K, 1080p or more for video. A lower setting is only used if you choose it explicitly. There is no "premium" plan that unlocks better output, and the cost in credits follows the real cost of the provider.

The models Miravela runs today, and what each one is best at.
Model	Maker	Type	Best at
Nano Banana 2	Google	Image	The editorial workhorse — fashion realism, skin, fabric, film grain. Top-3 in blind human votes.
GPT Image 2	OpenAI	Image	The graphic-design brain — layouts, typography, text that stays readable inside the image.
MAI Image 2.5	Microsoft	Image	#1 of the human-voted image arena — pure photorealism, faces that hold up to zoom.
Seedream	ByteDance	Image	Cinematic stills — film lighting, moody color grades, a whole story in one frame.
Flux Pro	BFL	Image	Product precision — packshots, materials, jewellery detail, e-commerce accuracy.
PixVerse C1	PixVerse	Video	Current video SOTA in blind votes — emotion and character performance that feels directed.
Veo 3.1	Google	Video	Realistic video with native sound — ambiance, dialogue, brand films.
Kling	Kuaishou	Video	High movement — runway walks, fabric in motion, sport, energy.
Grok Imagine 1.5	xAI	Video	Fast, stylized clips — social-first, punchy, made to stop the scroll.
Wan 2.7	Alibaba	Video	World building — sets, environments, wide establishing shots.
Seedance	ByteDance	Video	Fluid choreography — dance, gesture, rhythm-cut movement.
Hailuo	MiniMax	Video	Expressive faces — close-up emotion, beauty shots, subtle acting.

Questions people ask

Can I choose the model myself?: Yes. The agent proposes a model per job and says why; you can override it on any row before running. Autopilot, when you enable it, lets the agent decide and run by itself.
Do you add new models when they come out?: Yes, once they win their category in blind human votes and in our own production. A model can also be removed when it is discontinued by its maker or overtaken. The list on this page is the live one.
Do all models accept real faces and products as references?: Most do; some refuse real faces unless rights are verified. Miravela routes the job to a model that accepts your reference, or tells you why a model refused it, and you can use another one.