WaveSpeedAI review for AI personas: pay per image
WaveSpeedAI for AI persona work: pay-as-you-go prices per image and per second, the models it runs, LoRA training by API, and who should pick a subscription.
At a glance
Pricing
Pay as you go with no subscription: from $0.005 per image and $0.01 per second of video, plus $1 of free credits on signup, as read on its pricing and model pages on 5 October 2026.
This price comes from the vendor's own pricing page, on the date shown with it. These tools often rename their plans and change prices, so check before you pay.
Worth the money if
You generate in bursts. You want to try many models cheaply before you pick one. Or you are comfortable with an API and want to train your own LoRA and use it from a script.
Not for you if
You want a guided studio with presets and a built-in character tool instead of a long list of models. Or you would rather pay a fixed amount each month than keep an eye on a balance.
What WaveSpeedAI does
WaveSpeedAI is a catalogue of AI models with an API and a browser playground. It has image, video and language models from many vendors. You run each model on its own and pay per output, with no subscription. We have used it for persona work. How to use WaveSpeed AI shows the playground, the API key and the price checks step by step. This page only reports what WaveSpeed’s own pages say, and what that means if you run an AI persona.
WaveSpeedAI pricing: per image and per second
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. Some premium models may not be available with trial credits. Starting prices are list prices at the lowest resolution or quality: per image, Z-Image Turbo $0.005, Seedream 5.0 Pro $0.045 and Nano Banana 2 $0.045; per second of video, MiniMax H3 $0.04, Veo 3.1 Fast $0.10 and Seedance 2.0 $0.12. Volume discounts are available on request.
Look carefully at the unit. Images cost per image, but video costs per second. A ten-second clip at $0.10 a second costs a dollar, not ten cents. You can only compare two video models when both prices use the same unit. For that reason the model price table always shows the unit next to the price. WaveSpeed vs fal vs Replicate prices the same six models on all three hosts on one day.
This way of paying is simple, and it fits you if you generate in bursts. You lose nothing in a quiet month, and in a busy weekend you pay only for the two hundred images you make. The downside is that nothing is included. You have no fixed monthly amount to plan with, and no interface built to keep one character consistent. You can build that yourself: Claude Code image generation for an AI persona puts your reference set and a price check into one skill on top of WaveSpeed’s MCP server.
Account levels: top up once before a busy weekend
Every account starts at Bronze: 5 predictions per minute and 2 running at the same time. A single successful top-up below $1,000 moves the account to Silver: 500 predictions per minute and 300 at the same time. A single top-up of $1,000 to $4,999 gives Gold, and $5,000 or more gives Ultra. The level upgrades automatically after the top-up.
WaveSpeedAI docs — Account Levels & Rate Limits (opens in a new tab)checked
On the free $1 alone you are on Bronze. Five generations a minute, two at a time, is fine for a test, but it is slow for a batch of two hundred images. Your first top-up of any size under $1,000 moves you to Silver. Five hundred a minute is far more than one person needs, so you do not need Gold or Ultra.
Choosing a model for a persona
The big catalogue is the strength of the platform, and it is also what makes it hard to use. There are many models, and no vendor publishes how well a model keeps a face. For a persona, that decides whether you can use a model at all. So use the low per-image prices to test. Run the same reference set through two or three cheap models. Look at the angles where your character breaks. Only then move up to the expensive models. Which AI video models hold a face shows how to do that for video.
Training your own LoRA
The Flux Dev LoRA Trainer takes a URL to a zip archive of images, recommending at least 4 and more where possible, with an optional trigger word, training steps defaulting to 1,000 (500 to 10,000), a learning rate defaulting to 0.0004 and a LoRA rank defaulting to 16 (1 to 64).
WaveSpeedAI docs — Flux Dev LoRA Trainer (opens in a new tab)checked
For persona work, this makes WaveSpeed useful for more than cheap images. You can train a character LoRA on it and then generate with that LoRA, and you do not need your own GPU. Two of the defaults differ from this site’s own LoRA settings. First, four images is only the technical minimum. A face that stays the same across angles needs a proper reference set of twenty or more. Second, the default learning rate of 0.0004 is higher than the 1e-4 starting point the guide recommends. Set it yourself and do not just accept the default.
The API documentation still does not say what training costs, how long it takes, or whether you get the file. The trainer’s model page and WaveSpeed’s own guides do.
The Flux Dev LoRA Trainer starts at $1 per run. The final charge scales with the settings and is shown next to the button before you submit. WaveSpeed reports a generation time of around 405 seconds per request.
WaveSpeedAI — Flux-dev LoRA Trainer (opens in a new tab)checked
WaveSpeed’s training guide says 1,000 steps usually take around 8 minutes and 3,000 steps around 25 minutes, depending on the model and workload. After training you can download the LoRA file to your computer or copy the hosted model URL. WaveSpeed keeps the trained model available for 7 days and suggests uploading it to Hugging Face to keep it permanently.
WaveSpeedAI blog — How to train your own LoRA model without coding (opens in a new tab)checked
WaveSpeed’s guide to using a LoRA says LoRAs on WaveSpeedAI are typically loaded from a single .safetensors file, and that a LoRA must exactly match the base model: the same family, version and parameter size.
WaveSpeedAI blog — Use LoRA on WaveSpeedAI (opens in a new tab)checked
So you can take the LoRA with you, and that matters most. A downloaded LoRA works anywhere that runs the same base model. Download it right after training, because WaveSpeed only keeps it for 7 days. Remember the match rule too: a LoRA from the Flux Dev trainer works with Flux Dev, not with a video model such as MiniMax H3.
Check commercial use and provenance first
The pricing page says nothing about commercial use of outputs or about provenance metadata. For commercial use, read the terms. For provenance, check one exported file with a manifest inspector. The EU AI Act makes the generator’s maker put provenance markers in. It does not ban you from removing them. Keep them anyway: platforms read them to label your posts automatically, and a generator’s terms may forbid removing them. Don’t strip your watermarks shows how.
Alternatives
5,856+ members
In our Discord, people who build AI influencers help each other. It is free, and you can read everything right away.
Join the DiscordWhat is inside the server