Disclosure first, because this is a guide to our own thing: LU Labs is the hosted layer we built next to our open source desktop app for Windows and Linux, which is free and runs models locally. This post is about the hosted image side, what a render costs in credits, and where the hosted route is the wrong choice.
The hardware problem, briefly
Flux 2 Dev is a large model. Running it locally at a quality you would actually ship means a 24 GB card, and the workarounds below that line are the usual ones: heavy quantization, sequential CPU offload, or a tiling scheme that turns a render you would have watched into a render you go and make coffee during. All of them work. None of them are fun on a laptop.
The alternative is to keep the prompt and send the compute elsewhere. That is what this is.
The route
Sign in at lu-labs.ai and open the Studio in your browser. Nothing to install, no Python environment, no xformers version to reconcile. Every hosted account gets the same 10 image models:
- Flux 2 Dev, the one in the title
- Flux Dev and Flux Schnell, the older pair, Schnell being the fast draft model
- Qwen Image
- HiDream
- HunyuanImage 2.1
- Z-Image Turbo
- Chroma
- Prefect Pony XL
- Neta Lumina
The image catalog does not scale with the plan. It is identical on every hosted plan and identical on a pack, and the chat catalog works the same way, so the only thing price changes is how many credits you get.
Working, not just generating
A first render is rarely the last one, so the editing tools matter more than the model count:
- Inpainting on Flux Dev. Mask the part that is wrong, describe what should be there instead.
- Qwen Image Edit. Instruction-driven edits on an existing image rather than a fresh sample.
- Background removal and an eraser for the cleanup passes that otherwise send you back to a desktop editor.
- Upscale to 2k, 4k or 8k, which is the step people forget until they try to print something.
The workflow that gets the most out of a balance is boring and effective: draft wide on Flux Schnell until the composition is right, then re-render the keeper on Flux 2 Dev, then inpaint the one detail that is off, then upscale. Cheap exploration, expensive commitment, in that order.
What it costs, in whole numbers
Credits, one pool, one credit is $0.00001. Images are priced per render at a fixed rate per model:
| Model | Credits per image |
|---|---|
| Flux Schnell | 300 |
| Flux Dev | 1,200 |
| Flux 2 Dev | 1,200 |
That converts cleanly:
| Flux Schnell renders | Flux 2 Dev renders | |
|---|---|---|
| €5 pack, 165,000 credits | about 550 | 137 |
| Hosted month, 900,000 credits | about 3,000 | 750 |
The €5 pack is the entry point in the title and it is a real one. No subscription is required for it, the credits never expire, and you can buy any pack on the first purchase: €5 for 165,000 credits, €10 for 350,000, €25 for 900,000. If you render in bursts, a pack is the honest shape, because there is no month to waste.
The Hosted plan is €19 a month for 900,000 credits, and it adds two things a pack does not have: LoRA training, 2 a month, and a monthly refill you do not have to think about.
The full rate card is on the pricing page.
The same balance does video
Worth knowing before you decide how much to buy, because video is where a balance disappears fastest. Every hosted account also gets 5 video models: Wan 2.2 720p, Wan 2.2 Fast, LTX-2 with audio, LTX 2.3 and HunyuanVideo 1.5, with clips at 5 or 8 seconds. An LTX-2 clip at 5 seconds is 8,000 credits, a different order of cost from a still image. Plan for that, or keep video and stills on separate purchases.
Where the hosted route is wrong
You want a specific community checkpoint. The catalog is curated and there is no bring-your-own checkpoint. If your work depends on one particular merge from a model site, hosted will not do it and local will.
You want the render to never leave your machine. Hosted means the prompt reaches a GPU that is not yours. Accounts and billing sit in the EU, we do not train on user data and we never sell it, but that is a policy rather than physics. The free desktop app on Windows or Linux is the version where the guarantee is physical, and that is exactly why we shipped it first.
You render constantly. If a 24 GB card is already in the machine and it is busy every day, local is cheaper than any credit pool, and it always will be. Hosted wins on the days your hardware says no, on a laptop, or on a Mac where we have no desktop build and the browser Studio is the only route.
One more honest limit: the hardware is shared. It is managed H100, A100 and B200 class GPUs, which is more card than most people have, but at busy hours there is a short queue. A render that normally comes back quickly sometimes waits.
The short version
If you have the card, run Flux 2 Dev locally. If you do not, a €5 pack gets you 137 Flux 2 Dev renders or about 550 quick drafts, with no subscription and no expiry date on the credits, and the editing and upscale tools come with it rather than as a separate purchase.
Start with a Hosted month or a €5 pack and the longer guide is at how to generate Flux 2 images online.
Top comments (0)