Flux in Detail

[ikaijua]

Flux in Detail

Flux is Black Forest Labs' image family; the lineup is FLUX.2 — photorealistic, multi-reference editing, on-device open weights, API for commercial use.

What it is. By Black Forest Labs (Germany); first-tier image generation alongside GPT Image and Nano Banana.
Signature capability. Photorealism + up to 8-image reference editing; [klein] 4B is open-source and runs on-device.
Leaderboards. LMArena text-to-image overall: flux-2-max 1162±3; open-weight board: FLUX.2 [dev] #2.
Price. Per-megapixel: [pro] ~$0.03/MP; a 1024² image ~$0.03, no subscription floor.
Note. No official free tier; dev/9B open weights are non-commercial only.

Floors below follow “Product intro → Leaderboards & pricing → Capabilities & deployment → FAQ & resources”.

Table of Contents

Product intro

1. What is Flux?
2. Where to use it

Leaderboards & pricing

3. Leaderboards & hands-on
4. Pricing

Capabilities & deployment

5. Model family & core capabilities
6. On-device & open source
7. Comparison with other image models

FAQ & resources

8. FAQ
9. Sources

Clicking a TOC item jumps to the floor; if it does not respond in-page, open the link in a new tab.

[ikaijua]

What is Flux?

Flux is the image generation family from Black Forest Labs (Germany; founders from Stability AI). Current lineup: FLUX.2 (2026); FLUX 3 is next-gen multimodal.

Unlike closed models such as GPT Image and Nano Banana, Flux runs a dual track of commercial API + open weights:

• Flagship variants: API-paid
• [dev] and [klein] variants: open weights
• [klein] 4B: can even run on-device

[ikaijua]

Where to use it

Entry	Notes	Link
BFL official API	Docs and pricing (docs.bfl.ml)	https://docs.bfl.ml
BFL web (bfl.ai)	Official product page; FLUX 3 live	https://bfl.ai
fal.ai	Popular inference host, pay-per-megapixel	https://fal.ai/flux-2
Replicate	Cloud inference; FLUX.2 [pro] from ~$0.015/image	https://replicate.com
DeepInfra	Per-image billing; klein 4B from $0.014	https://deepinfra.com/flux
OpenRouter	Aggregated API; image model comparison entry	https://openrouter.ai

Most platforms offer limited free trial credits; for frequent high-volume use, compare per-image pricing before committing.

[ikaijua]

Leaderboards & hands-on

LMArena Text-to-Image overall (updated 2026-09, Elo higher is better):

• flux-2-max: 1162±3 (front of the pack, competing directly with the Nano Banana line)
• flux-2-klein-9b: 1070±… (among the leading open-weight entries)

Category boards:

• Commercial design: flux-2-flex 1173±5
• 3D modeling: flux-2-flex 1173±7, flux-2-max 1161±7
• Open-weight board: FLUX.2 [dev] ranked Update README.md #2 (The Open Weights)

Community takes:

• Photorealistic portraits/texture: reviewers rate FLUX.2 Pro a match for Nano Banana 2 (some comparisons put it ahead)
• Image editing: FLUX.2 [pro] merges up to 8 reference images (10 in playground) with strong edit consistency
• vs GPT Image 2: Flux wins on final-art texture/material realism; GPT Image is steadier on Chinese prompts and in-image text

Caveat: leaderboard scores and official claims have different methodologies — judge by the images you actually generate.

[ikaijua]

Pricing

Billed per megapixel (MP), no subscription floor (fal.ai / BFL API):

Variant	Price	Positioning
FLUX.2 [klein] 4B	from ~$0.014/image	Real-time, high volume
FLUX.2 [klein] 9B	from ~$0.015/image	Balanced quality/speed
FLUX.2 [dev]	from $0.012/MP	Open weights, best value
FLUX.2 [pro]	$0.03/MP (edit from $0.045)	Production-grade
FLUX.2 [flex]	$0.05/MP	Typography & fine detail
FLUX.2 [max]	$0.07/MP	Maximum quality

A standard 1024x1024 image costs ~$0.03 — cheaper at scale than GPT Image 2.5 (token-billed, ~$0.21/image at high tier).

[ikaijua]

Model family & core capabilities

FLUX.2 family (on sale in 2026):

• [max]: performance ceiling. Highest edit consistency, strongest prompt following, grounded generation with real-time web context
• [pro]: production-grade value pick. Up to 8-image reference editing, photorealistic output
• [flex]: specialized in typography and small-detail preservation
• [klein] 4B / 9B: small on-device models built for speed
• [dev]: open weights (non-commercial license)

Core capabilities:

• Multi-reference editing: merge up to 8 input images (10 in playground) — e.g. "apply design from image 1 onto the cup in image 2"
• Photorealistic texture: narrows the gap between generated and real imagery
• Chinese prompt support (klein onward)
• Kontext line: generation grounded in real-time web context

FLUX 3 (Early Access 2026-07-23): unified multimodal foundation model for image + video + audio; video up to 20 seconds with native audio and multilingual dialogue, video GA 2026-08-04. Still early-stage — playable, but aimed at developers.

[ikaijua]

On-device & open source

• FLUX.2 [klein] 4B: Apache 2.0 open source; runs on 8GB VRAM laptops with a sub-5-second (target sub-second) generation goal; preloaded in ASUS ProArt (RTX GPU) laptops via the MuseTree app
• FLUX.2 [klein] 9B: open weights, non-commercial license (flux-non-commercial-license)
• FLUX.2 [dev]: open weights, free for personal/research use
• Ecosystem: rich ComfyUI workflows (text-to-image, editing, outpainting, inpainting all have ready nodes); official inference repo black-forest-labs/flux2

To run locally: an 8GB+ VRAM GPU is enough to start; for commercial output, prefer the API.

[ikaijua]

Comparison with other image models

Dimension	Flux (FLUX.2)	GPT Image 2.5	Nano Banana 2
Maker	Black Forest Labs	OpenAI	Google
Model form	Commercial API + open weights	Closed API	Closed (default in Gemini app)
Edge	Photorealism, multi-reference editing, on-device	Precise local editing, in-image Chinese text	Speed, 4K, character consistency
Price (per image ref)	~$0.03 (pro, 1MP)	~$0.21 (high, 1024²)	~$0.067 API (1K)

Picking one: Flux for photorealism / batch / local; GPT Image for in-image text; Nano Banana for speed and character consistency.

[ikaijua]

FAQ

Can free users use Flux?
No official permanent free tier. Try each platform's trial credits, or run the open-weight variants (dev / klein) locally.

Which variant should I pick?
[klein] or [dev] for high-volume everyday work; [pro] for production-grade editing; [max] for maximum quality and edit consistency; [flex] for pure typography.

What is FLUX 3?
A multimodal foundation model (image + video + audio), EA July 2026, video GA August 2026; still early and mostly for developers.

Local deployment requirements?
klein 4B runs on 8GB VRAM (Apache 2.0); dev/9B need more VRAM and are non-commercial only.

Is it privacy-safe?
API generations follow each host's terms; for privacy-sensitive work, run klein locally, fully offline.

[ikaijua]

Sources

• Black Forest Labs: https://bfl.ai
• Official release notes: https://docs.bfl.ml/release-notes
• Official pricing: https://docs.bfl.ml/quick_start/pricing
• FLUX 3 announcement: https://bfl.ai/blog/flux-3
• LMArena text-to-image board: https://arena.ai/leaderboard/text-to-image/overall
• fal.ai FLUX.2 page: https://fal.ai/flux-2
• Official inference repo: https://github.com/black-forest-labs/flux2
• Chinese version of this post: Flux in Detail (Chinese)