Flux 2 Dev vs Flux Dev vs Flux Schnell: Which One for Which Job

The Flux family has grown into a naming problem. Schnell, Dev, and now Flux 2 Dev, all three in the hosted image catalog, all three sitting in the same dropdown, and nothing in the interface telling you which one your current task wants.

The useful way to separate them is not by quality rankings, which are subjective and change with the prompt. It is by the two things that are objectively different: what a render costs, and what each one can do that the others cannot.

The cost line, which is the sharpest difference

Flux Schnell is 300 credits per image. Flux Dev and Flux 2 Dev are 1,200 credits each. One credit is $0.00001.

That ratio is four to one, and it is the single most important fact about choosing between them. Against a €5 credit pack of 165,000 credits, it means about 550 images on Schnell or about 137 on either Dev model. On a Hosted month at €19 with 900,000 credits, about 3,000 or about 750. The plans and packs are here.

Notice what that does to a working session. If you render a run of variations to find a composition and then render the keeper, doing all of it on Flux 2 Dev costs you the same as doing the exploration on Schnell and the keeper on Flux 2 Dev, four times over. The exploration is where the volume is, and the exploration is where the quality matters least.

Flux Schnell: the drafting model

Schnell is built for speed, and speed is what a first pass needs. You are not evaluating fine texture on an early pass. You are asking whether the subject sits well in the frame, whether the light direction reads, whether the palette is anywhere near what you meant.

Use it for:

• Finding a composition before you commit.
• Testing whether a prompt structure works at all.
• Bulk variations where you will keep one and discard the rest.
• Anything going into a mood board rather than into a deliverable.

The mistake is treating Schnell as the budget option you settle for. It is not a compromise, it is a different stage of the same job.

Flux Dev: the one with the edit lane

Flux Dev is the quality model of the earlier generation, and in our catalog it carries something the others do not: masked editing. Paint over a region, describe only that region, and the rest of the frame stays exactly as it was.

That makes Flux Dev the answer whenever the picture is nearly right. One hand is wrong, one object in the background does not belong, one label reads as gibberish. Re-rendering the whole frame to fix a corner throws away everything that worked and rolls the dice again on all of it.

So Flux Dev is less "the older Dev" and more "the repair model". Even in a workflow where Flux 2 Dev produces the frames, Flux Dev is where the frames get fixed.

The other repair tools sit alongside it: Qwen Image Edit for instruction based changes to the whole frame, background removal, an object eraser, and upscaling at 2k, 4k and 8k.

Flux 2 Dev: the finished frame

Flux 2 Dev is the current generation quality model, same 1,200 credits as Flux Dev, and it is what you point at the render you intend to keep. If you only ever use one model from the family for output, it is this one.

Use it for:

• The final version of a composition you already validated on Schnell.
• Anything that leaves your machine: a client, a site, a print.
• Prompts with a lot of clauses, where the model has to reconcile several instructions at once.

Because it costs the same as Flux Dev, there is no reason to draft on Flux Dev. Draft on Schnell, finish on Flux 2 Dev, repair on Flux Dev.

The workflow that falls out of this

Three models, three stages, and the order is not negotiable if you care about the bill:

1. Schnell until the composition is right.
2. Flux 2 Dev for the keeper.
3. Flux Dev masked editing for the one thing that is still wrong, then upscale once at the end.

Upscaling a draft is the classic waste. So is running the exploration phase on a quality model because it "might as well be good".

What about the other seven

The Flux three are not the whole catalog. There are 10 hosted image models in total, and the rest are there because they are genuinely different rather than incrementally better: Qwen Image, HiDream, HunyuanImage 2.1, Z-Image Turbo, Chroma, Prefect Pony XL and Neta Lumina. Different training data means different default aesthetics, and for stylised work one of those may get you closer in one shot than a paragraph of prompt engineering on Flux ever will.

If you need the same character or the same object across many frames, none of the three Flux models will do it by prompting alone. That is what LoRA training is for, and Hosted includes 2 trainings a month. A credit pack on its own includes none.

The practical next step

All three are in the same picker on the entry plan, so this is not a purchasing decision between models. It is a habit you build inside one account. The background on hosted rendering is on the FLUX 2 online without a GPU page, the working procedure is in how to generate Flux 2 images online, and if you would rather run the whole thing on your own card, generating images on your own computer covers that setup instead.

Rendering runs on managed NVIDIA H100, A100 and B200 class GPUs, so a job can queue for a moment at peak. Credits are one pool for text, images and video, the catalog is curated so you cannot bring your own base checkpoint, and hosted image and video follow the Terms of Service. Content involving minors is refused everywhere, always. Account data is hosted in the EU and you can delete it yourself.

If you take one rule from this: never explore on a 1,200 credit model. Start on Hosted and put Schnell first in your habit, not last.