Skip to main content

FLUX vs SDXL: licence, VRAM and price per image

FLUX is now a family. Black Forest Labs sells FLUX 3 Image and FLUX.2 through its API, publishes FLUX.2 [klein] and FLUX.2 [dev] as open weights, and lists FLUX.1 as its previous generation. Stable Diffusion XL is one fixed model: SDXL 1.0, released by Stability AI on 26 July 2023 and still the base of thousands of LoRAs and fine-tunes.

Reviewed against 36 sources. Prices, models, policies and terms change, so check the linked pages before you buy, build, publish or trade.

The short answer

Start new work on FLUX. FLUX.2 [klein] 4B is Apache 2.0, edits as well as generates, and every FLUX model on the Artificial Analysis leaderboard scores above SDXL 1.0 in blind votes. Stay on SDXL if your graphics card has 8 GB of VRAM, if your workflow is built on SDXL LoRAs and ControlNets, or if $0.009 an image matters more than quality.

Choose FLUX if

  • You want a model that is still developed. FLUX 3 Image launched on 1 October 2026; the last commit to the SDXL 1.0 repository is dated 30 October 2023.
  • You need readable text or edits from reference images. FLUX.2 [pro] takes up to 8 references per API request; the SDXL model card says it cannot render legible text.
  • You want commercial use with no licence fee. FLUX.2 [klein] 4B is Apache 2.0, and its model card says about 13 GB of VRAM.

Choose SDXL if

  • Your graphics card has 8 GB of VRAM, the size Stability AI names for SDXL 1.0. The model card of the smallest FLUX.2 model says about 13 GB.
  • Your workflow runs on SDXL checkpoints, LoRAs and ControlNets. None of them load into FLUX.
  • You want the cheapest image from the maker's API: $0.009 on SDXL 1.0 against $0.014 on FLUX.2 [klein] 4B.

The facts, side by side

Each row shows the date it was checked and numbered links to the pages it was read from.

marks the side the facts favour on that row. Rows without one are a draw or a matter of fit.

Swipe the table sideways to see SDXL.

FLUX and SDXL compared, fact by fact
ComparedFLUXBlack Forest LabsSDXLStability AI
Who makes and hosts it1162127Black Forest Labs (BFL). Hosted on the BFL API; open weights on Hugging Face.Stability AI. Hosted on the Stability AI API; open weights on Hugging Face.
What the name covers142127A family: FLUX 3 Image, FLUX.2 [max], [pro], [flex], [klein] and [dev], and FLUX.1, which BFL lists as its previous generation.One model, SDXL 1.0: a base checkpoint with an optional refiner. The Stability AI price list calls it a "legacy base model".
Released14324FLUX.2 [dev]: 25 November 2025. FLUX.2 [klein]: 15 January 2026. FLUX 3 Image: 1 October 2026.SDXL 1.0: 26 July 2023
Model size1617191824FLUX.2 [klein]: 4 billion or 9 billion parameters. FLUX.1 [dev]: 12 billion. FLUX.2 [dev]: 32 billion.3.5 billion parameters in the base model, 6.6 billion with the refiner
Licence on the open weights16201121Apache 2.0 on FLUX.2 [klein] 4B and FLUX.1 [schnell]. FLUX Non-Commercial License on FLUX.2 [klein] 9B, FLUX.2 [dev] and FLUX.1 [dev].CreativeML Open RAIL++-M on the whole model
Commercial use on your own hardware111022Free on the two Apache 2.0 models. The others need a paid Commercial Weights License (Builder, Platform, Professional or Enterprise).Advantage: Allowed at no charge, including hosting it for others, if you pass on the use restrictions in Attachment A.
VRAM to run it1614171524FLUX.2 [klein] 4B: about 13 GB on its model card, about 8 GB in the BFL GitHub README. [klein] 9B: about 29 GB. FLUX.2 [dev]: an H100-class GPU, or about 18 to 20 GB with 4-bit quantization.Advantage: About 8 GB: Stability AI says SDXL 1.0 works on consumer GPUs of that size.
LoRA support326191821Yes. Diffusers has FLUX.1 and FLUX.2 loaders, and BFL documents LoRA training on FLUX.2 [klein] Base. Hugging Face lists 42,697 adapters on FLUX.1 [dev] and 80 on FLUX.2 [dev].Yes. Diffusers has a loader that works only with SDXL. Hugging Face lists 9,750 adapters on SDXL base 1.0.
ControlNet support30333531834FLUX.1: Canny and Depth control models from BFL, plus community ControlNets from InstantX, Shakker Labs and XLabs in Diffusers and ComfyUI. FLUX.2: no ControlNet pipeline on the Diffusers page; pose and layout come from reference images.Advantage: Yes. Diffusers has an SDXL ControlNet pipeline and training script, and warns that many SDXL ControlNet checkpoints are experimental.
Text inside images521Advantage: BFL describes FLUX.2 [flex] as specialized for typography and text rendering.The model card lists "cannot render legible text" as a limitation.
Still being updated32325Advantage: Yes. FLUX.2 [klein] launched on 15 January 2026 and FLUX 3 Image on 1 October 2026.No new weights. Last repository commit: 30 October 2023. Stability AI now leads its image page with Stable Diffusion 3.5.
Blind preference Elo (Artificial Analysis)36Advantage: FLUX 3 Image: 1116. FLUX.2 [dev]: 1000. FLUX.2 [klein] 4B: 863. FLUX.1 [dev]: 841. FLUX.1 [schnell]: 803.SDXL 1.0: 678, ranked 164 of the 168 models listed

Where each one is stronger

FLUX Black Forest Labs

  • New models keep arriving

    BFL released FLUX.2 [dev] on 25 November 2025, FLUX.2 [klein] on 15 January 2026 and FLUX 3 Image on 1 October 2026. The SDXL 1.0 repository has not changed since 30 October 2023. 14323

  • Preferred in blind tests

    Artificial Analysis ranks image models by Elo from blind pairwise votes. On its AA-Image-T2I v2.0 board, read on 13 October 2026, FLUX.2 [klein] 4B scores 863 and FLUX.1 [dev] 841, against 678 for SDXL 1.0. 36

  • Text, colour and references in one model

    FLUX.2 generates and edits with the same model, accepts hex colour codes and takes up to 8 reference images per API request, or 4 on [klein]. The SDXL model card lists legible text and compositional prompts as limitations. 521

SDXL Stability AI

  • One licence, commercial use included

    All of SDXL 1.0 is under CreativeML Open RAIL++-M, a no-charge, royalty-free grant that covers fine-tuning, distribution and hosting, with a list of restricted uses. On the FLUX side only FLUX.2 [klein] 4B and FLUX.1 [schnell] are that open. 2211

  • Runs on an 8 GB card

    Stability AI says SDXL 1.0 should work on consumer GPUs with 8 GB of VRAM. The FLUX.2 [klein] 4B model card says about 13 GB. 2416

  • A documented ControlNet and LoRA stack

    Diffusers documents a ControlNet pipeline and training script for SDXL. Hugging Face shows 9,750 adapters, 1,225 fine-tunes and 4,300,160 downloads in the last month for SDXL base 1.0. 3421

Which to pick, by job

  • Keeping an AI character consistent across social posts

    Pick FLUX

    FLUX.2 [pro] takes up to 8 reference images per API request and [klein] up to 4, so a face and an outfit carry across scenes without training. For a trained look, BFL lists 12 GB of VRAM to train a LoRA on [klein] Base 4B. 56

  • A laptop or desktop with an 8 GB graphics card

    Pick SDXL

    Stability AI names 8 GB of VRAM for SDXL 1.0. BFL gives about 13 GB for FLUX.2 [klein] 4B on its model card and about 8 GB in its GitHub README, so test [klein] 4B but plan on SDXL. 241614

  • Posters, thumbnails and product shots with readable text

    Pick FLUX

    BFL describes FLUX.2 [flex] as specialized for typography, and FLUX.2 accepts hex codes for exact brand colours. The SDXL model card says the model cannot render legible text. 521

  • A pipeline already built on SDXL LoRAs and ControlNets

    Pick SDXL

    SDXL LoRAs load into the SDXL UNet and FLUX LoRAs into a transformer, so nothing carries over: every LoRA is retrained and every ControlNet replaced. Stay on SDXL until a FLUX model beats your results on your own prompts. 323433

  • A paid product that generates images on your own GPU

    Pick FLUX

    FLUX.2 [klein] 4B is Apache 2.0, so commercial use needs no licence, and it scores 863 against 678 for SDXL 1.0 in Artificial Analysis blind votes. FLUX.2 [dev], [klein] 9B and FLUX.1 [dev] need a paid Commercial Weights License first. 161136

What it costs

List prices in US dollars for one image, read from each maker's own pages. Both bill in credits worth $0.01. FLUX prices are starting prices at about 1 megapixel and rise with output size; the SDXL price rises with steps. Other hosts set their own prices and are not listed.

marks the side the facts favour on that row. Rows without one are a draw or a matter of fit.

Swipe the table sideways to see SDXL.

FLUX and SDXL prices
ComparedFLUXBlack Forest LabsSDXLStability AI
Downloading the weights161817192021$0. FLUX.2 [klein] 4B has no gate; FLUX.2 [dev], [klein] 9B and both FLUX.1 models ask for contact details first.$0, with no gate
Commercial licence for your own hardware11101222$0 on FLUX.2 [klein] 4B and FLUX.1 [schnell]. For the rest, the Builder tier (FLUX.2 [klein] models, 10K images a month, 1 domain) is billed monthly in the BFL dashboard; its price is not published on the public pages.Advantage: $0. The licence is a no-charge, royalty-free grant.
Cheapest image on the maker's API12728FLUX.2 [klein] 4B: from $0.014Advantage: SDXL 1.0: $0.009 (0.9 credits) at 1024x1024 with 30 steps or fewer
How the price rises128With size: FLUX.2 [klein] 4B adds $0.001 per extra megapixel, so 2 megapixels is $0.015.With steps: above 30, the cost is 0.9 credits times steps divided by 30, so 50 steps is $0.015.
Main production model12729FLUX.2 [pro]: from $0.03 for text to image, from $0.045 for editingStable Image Core, which Stability AI calls the next generation after SDXL: $0.03 (3 credits)
Highest quality tier127FLUX.2 [max]: from $0.07. FLUX 3 Image: $0.048 at 1k, $0.100 at 2k, $0.607 at 4k.No higher SDXL tier. Stable Diffusion 3.5 Large is $0.065 and Stable Image Ultra is $0.08.
Your own LoRA on the maker's API7128Advantage: Public beta on FLUX.2 [klein], billed at the base rate, from $0.014. One LoRA per request.Not offered. The SDXL 1.0 endpoint takes prompts, size, steps, sampler, seed and a style preset.
Free credits227None listed on the pricing page25 at sign-up, enough for 27 SDXL 1.0 images
1,000 images at 1024x102412736$14 on FLUX.2 [klein] 4B, $30 on FLUX.2 [pro], $48 on FLUX 3 ImageAdvantage: $9 on SDXL 1.0 at 30 steps or fewer

What changed recently

  1. Black Forest Labs released FLUX 3 Image on its API, priced from $0.041 to $0.607 an image by resolution. 3

  2. Black Forest Labs announced FLUX 3. Its launch plan lists open-weight access to a multimodal backbone, FLUX 3 Dev, as a later step. 9

  3. Stability AI updated its Core Models list. It names SDXL Turbo and Stable Diffusion 3.5 but not SDXL 1.0, which stays under its own licence. 26

  4. Black Forest Labs opened LoRA inference on FLUX.2 [klein] in public beta, billed at the base endpoint rate. 3

  5. Black Forest Labs launched FLUX.2 [klein]: 4B under Apache 2.0 from $0.014 an image, 9B under the FLUX Non-Commercial License. 3

  6. Black Forest Labs released FLUX.2 [dev], a 32 billion parameter open-weight model. 14

  7. Stability AI retired its Stable Diffusion 1.6 API and told those users to move to SDXL. 29

  8. Last commit to the SDXL base 1.0 repository on Hugging Face. 23

Questions people ask

Is FLUX better than SDXL?

In blind preference tests, yes. On the Artificial Analysis text to image leaderboard, SDXL 1.0 has an Elo of 678 and ranks 164 of the 168 models listed, while FLUX.1 [dev] has 841, FLUX.2 [klein] 4B has 863 and FLUX 3 Image has 1116. The votes compare base models, not community fine-tunes, so run your own prompts if you rely on a tuned SDXL checkpoint. 36

Can I use FLUX and SDXL commercially?

SDXL 1.0, yes: its CreativeML Open RAIL++-M licence is a no-charge grant to use, fine-tune, distribute and host the model, with a list of restricted uses. FLUX, partly. BFL says API usage includes commercial rights, and FLUX.2 [klein] 4B and FLUX.1 [schnell] are Apache 2.0. FLUX.2 [dev], FLUX.2 [klein] 9B and FLUX.1 [dev] are under the FLUX Non-Commercial License: it says outputs may be used commercially, but running the model for revenue or in production needs a paid Commercial Weights License. 2211162013

How much VRAM do FLUX and SDXL need?

SDXL 1.0 needs about 8 GB, according to Stability AI. FLUX changes with the model. BFL gives about 13 GB for FLUX.2 [klein] 4B on its model card and about 8 GB in its GitHub README, about 29 GB for [klein] 9B, and an H100-class GPU for FLUX.2 [dev], or about 18 to 20 GB with 4-bit quantization. The Diffusers docs put FLUX.1 at about 50 GB of RAM and VRAM to load every component, with an FP8 recipe that runs in under 16 GB. 241614171530

Do SDXL LoRAs and ControlNets work with FLUX?

No. The Diffusers SDXL loader puts LoRA layers into the SDXL UNet and works only with SDXL; the FLUX loaders put them into a transformer. BFL is as strict inside its own family: a LoRA uploaded for FLUX.2 [klein] 9B fails on the [klein] 4B endpoint. ControlNets are separate as well: Diffusers uses a different ControlNet class for SDXL and for FLUX.1. 3273433

How much does one image cost through the API?

SDXL 1.0 costs $0.009 on the Stability AI API at 1024x1024 with 30 steps or fewer. On the BFL API, FLUX.2 [klein] 4B starts at $0.014, FLUX.2 [pro] at $0.03 and FLUX.2 [max] at $0.07, and FLUX 3 Image is $0.048 at 1k. BFL lists FLUX.2 [dev] as local only, with no hosted API. 27281

Is SDXL still being updated?

Not with new weights. The last commit to the SDXL base 1.0 repository is dated 30 October 2023, the Stability AI price list calls it a legacy base model, and the image page leads with Stable Diffusion 3.5. Stability AI still hosts SDXL 1.0 and in July 2025 named it as the replacement for its retired Stable Diffusion 1.6 API. 23272529

Sources

Each page was read on the date beside it. The links open the current version, which may have changed since.

  1. 1Black Forest Labs: API pricing
  2. 2Black Forest Labs: pricing page and calculator
  3. 3Black Forest Labs: release notes
  4. 4Black Forest Labs: models
  5. 5Black Forest Labs: FLUX.2 overview
  6. 6Black Forest Labs: FLUX.2 [klein] training
  7. 7Black Forest Labs: FLUX.2 LoRA inference
  8. 8Black Forest Labs: pose and layout guidance
  9. 9Black Forest Labs: FLUX 3 announcement
  10. 10Black Forest Labs: open weights licensing
  11. 11Black Forest Labs: how FLUX licensing works
  12. 12Black Forest Labs: licensing billing and payment
  13. 13Black Forest Labs: FLUX [dev] Non-Commercial License v2.0
  14. 14Black Forest Labs: FLUX.2 repository on GitHub
  15. 15Black Forest Labs: running FLUX.2 [dev] with Diffusers
  16. 16Hugging Face: FLUX.2 [klein] 4B model card
  17. 17Hugging Face: FLUX.2 [klein] 9B model card
  18. 18Hugging Face: FLUX.2 [dev] model card
  19. 19Hugging Face: FLUX.1 [dev] model card
  20. 20Hugging Face: FLUX.1 [schnell] model card
  21. 21Hugging Face: SDXL base 1.0 model card
  22. 22Hugging Face: SDXL 1.0 licence (CreativeML Open RAIL++-M)
  23. 23Hugging Face: SDXL base 1.0 commit history
  24. 24Stability AI: SDXL 1.0 announcement
  25. 25Stability AI: image models
  26. 26Stability AI: Core Models list
  27. 27Stability AI: Developer Platform pricing
  28. 28Stability AI: API reference, SDXL 1.0
  29. 29Stability AI: Developer Platform release notes
  30. 30Hugging Face Diffusers: Flux pipeline
  31. 31Hugging Face Diffusers: Flux2 pipeline
  32. 32Hugging Face Diffusers: LoRA loaders
  33. 33Hugging Face Diffusers: ControlNet with Flux.1
  34. 34Hugging Face Diffusers: ControlNet with Stable Diffusion XL
  35. 35ComfyUI: Flux examples
  36. 36Artificial Analysis: text to image leaderboard (AA-Image-T2I v2.0)

Take the free Starter Kit.

Five working files from the IImagined programs: 32 Instagram hooks, an AI character template, a pricing worksheet and two calculators. Free, and they open in your browser.

Get the free Starter Kit

Related program

AI Influencers. Learn reference testing, model selection, quality control, disclosure, rights checks, production, and measurement. Included in All Access.

IImagined publishes this comparison and sells the All Access membership.