How to Generate Flux 2 Images Online

Diffusion models are the part of local AI that punishes hardware the hardest. A card with real VRAM, a working driver stack, tens of gigabytes of checkpoints on disk, and an evening of setup before the first picture. If you own that rig, use it, and our guide to generating images on your own computer covers the setup properly.

This guide is the other route: Flux 2 Dev rendering on managed GPUs with the finished image landing in your browser. The overview lives on the FLUX 2 online without a GPU page. What follows is the working procedure, including the part most guides skip, which is what each button costs.

Step 1: get credits, and know what they buy

There are two ways in and only one is a subscription.

A credit pack needs no plan at all: €5 for 165,000 credits, €10 for 350,000, €25 for 900,000. Any pack is available from your first purchase and the credits never expire.

Hosted is €19 a month with 900,000 credits, plus the chat, code and video studios on the same wallet.

In images that works out roughly like this. Flux Schnell is 300 credits per image. Flux Dev and Flux 2 Dev are 1,200 credits each. So a €5 pack is about 550 quick images, or about 137 on Flux 2 Dev. A Hosted month is about 3,000 quick images, or about 750 on Flux 2 Dev. One credit is $0.00001, if you want to check the arithmetic yourself. The full table is on the pricing page.

Those are estimates from one shared budget, not a quota. Chat, code and video draw from the same credits, and a 5 s LTX-2 clip is 8,000 credits, so a week of video eats a month of images.

Start on Hosted if you want the monthly budget, or buy a pack from the credits tab. Card details go to Stripe and the payment is secured with 3D Secure.

Step 2: open the Image studio and choose a model

There are 10 hosted image models: Flux 2 Dev, Flux Dev, Flux Schnell, Qwen Image, HiDream, HunyuanImage 2.1, Z-Image Turbo, Chroma, Prefect Pony XL and Neta Lumina.

Pick Flux 2 Dev when you want the finished frame. Which of the three Flux models suits which job is the whole subject of Flux 2 Dev vs Flux Dev vs Flux Schnell, and it is worth reading before you spend much.

Step 3: draft cheap, finish expensive

This is the habit that decides your bill. A Flux 2 Dev render costs four times what a Flux Schnell render costs, and most of your early renders exist to find out whether the composition works at all.

So: iterate on Flux Schnell. Change the prompt, look, change it again. When the framing, the subject placement and the general mood are right, move the same prompt to Flux 2 Dev for the version you keep.

People who skip this step burn a pack in an afternoon and conclude the pricing is steep. People who do it get a lot of finished frames out of the same money.

Step 4: write the prompt in layers

Flux models respond well to prompts that read like a shot description rather than a pile of adjectives. A structure that works:

A weathered brass diving helmet sitting on a workshop bench, lit from a
window on the left, dust visible in the light. Shallow depth of field,
the far end of the bench out of focus. Muted greens and oxidised copper.
Photographic, natural lens, no text anywhere in the frame.

Subject, then light, then depth, then palette, then medium. Say what should not be there, because "no text anywhere in the frame" catches a failure mode these models fall into on signage and labels.

Step 5: fix the part that is wrong

You rarely need a full re-render. The studio has repair tools that cost less attention and less time than starting over.

Masked editing on Flux Dev. Paint over the region you want changed, describe just that region, and the rest of the frame stays put. This is the right tool when one hand, one label or one background object is wrong.

Qwen Image Edit. Instruction based rather than mask based. You describe the change in a sentence and it reworks the frame. Better when the change is a property of the whole image rather than a patch of it.

Background removal and the object eraser. Cut the subject out, or delete something the model insisted on adding.

Step 6: upscale last

Upscaling is available at 2k, 4k and 8k. Do it at the end, on the frame you have already decided to keep. Upscaling a draft you are going to discard is the second most common way to waste credits after skipping step 3.

Step 7: train a character if you need consistency

If you need the same face or the same object across many images, prompt engineering will not get you there reliably. LoRA training will. Hosted includes 2 LoRA trainings a month. A credit pack on its own includes none, so if character consistency is the point of your project, that is the reason to take the plan rather than the pack.

You train on your own reference images and generate with the result afterwards. You cannot upload your own base checkpoint; the catalog is curated, and LoRAs sit on top of the models we host.

Troubleshooting

The render sits in a queue. Shared hardware. At busy hours a job waits before it starts. It clears on its own, and it does not cost extra.

The image ignores half my prompt. Cut the prompt down. Long prompts with many competing instructions tend to lose the ones in the middle. Get the core right on Flux Schnell first, then add detail one clause at a time.

Text in the image comes out as nonsense. Normal for diffusion models. Either add "no text anywhere in the frame" and put real text on afterwards in an editor, or accept it as texture rather than lettering.

My credits dropped faster than the image count suggested. Check what else ran. Video is the expensive one, and chat, code, image and video all share the wallet.

The masked edit changed things outside the mask. Tighten the mask and describe only what is inside it. Describing the whole scene in a masked edit invites the model to reconsider the whole scene.

A render was refused. Hosted image and video follow the Terms of Service. Content involving minors is refused everywhere, always.

The fine print worth knowing

Account data is hosted in the EU and you can delete it yourself. We never train on your data and we never sell it. Rendering happens on managed NVIDIA H100, A100 and B200 class GPUs, which is why none of this needs a card on your desk, and why the whole thing works the same from a phone as from a workstation.

Start on Flux Schnell, move to Flux 2 Dev when the composition is right, upscale once.