Flux 2 and Flux Kontext Explained: Every Variant, License, and Price (August 2026)
Last updated August 7, 2026
Flux 2 is Black Forest Labs' image family: one checkpoint that creates and edits at up to 4MP with 10 reference images, hex-color control, and JSON prompting. It trails GPT Image 2 on arena Elo (1196 vs 1339, Aug 2026) but leads every open-weights size class. Only klein-4B is Apache 2.0; dev-32B needs a paid license to deploy. All variants run inside the invideo agent.
Updated August 2026
Flux 2 is the current image-generation family from Black Forest Labs, the ~70-person German lab founded by ex-Stability researchers behind the original Stable Diffusion work. As of August 2026 the family spans a flagship 32B open-weights checkpoint, API-only pro/flex/max tiers, the sub-second Klein models, and the Kontext editing line — one architecture that both creates images and edits them from plain-language instructions, at up to 4MP with as many as 10 reference images per request. It is not the raw quality leader on blind arenas, but it is the strongest open-weights option in every size class it ships, and its per-megapixel pricing undercuts most closed rivals. All of it runs inside the invideo agent.
How did Flux get from FLUX.1 to Flux 2?
The family has moved fast since launch. Dates below are from Black Forest Labs' announcements:
- August 2024 — FLUX.1. Three checkpoints: schnell (Apache 2.0), dev (non-commercial), and API-only pro.
- October–November 2024 — FLUX 1.1 pro, then Ultra mode (4MP) and Raw mode for photorealism.
- November 2024 — FLUX.1 Tools: Fill (inpainting), Depth, Canny, and Redux control modules.
- May 29, 2025 — FLUX.1 Kontext pro and max: in-context editing driven by text instructions instead of masks. An open 12B Kontext dev followed in June 2025.
- July 2025 — FLUX.1 Krea dev, an open aesthetics-tuned collaboration checkpoint.
- November 25, 2025 — FLUX.2: pro and flex API tiers plus an open 32B dev checkpoint, all handling creation and editing in one model.
- December 16, 2025 — FLUX.2 max, adding grounding web search and up to 10 reference images.
- January 15, 2026 — FLUX.2 klein in 4B and 9B sizes, built for sub-second generation.
- 2026 Tools: Outpainting at 4MP, Erase, and Virtual Try-On (v1, then v2 with 4MP output and up to four layered garments).
In June 2026 the lab announced Martin Scorsese had joined as an advisor — a signal of where Black Forest Labs thinks its models sit in professional production.
Which Flux 2 variant should you pick?
invideo's model picker exposes the full family. One structural point first: "Flux 2 Pro Create" and "Flux 2 Pro Edit" are the same model. FLUX.2 uses a single checkpoint for generation and editing; the Create/Edit split is a mode (and price) distinction, not two models. The same logic applies to the non-pro Create/Edit pair.
| Picker entry | What it is | Best for | Official price (Aug 2026) |
|---|---|---|---|
| Flux 2 Pro Create | FLUX.2 [pro], generation mode | Default high-quality text-to-image | $0.03/MP |
| Flux 2 Pro Edit | FLUX.2 [pro], editing mode (same checkpoint) | Instruction-based edits of existing images | $0.045/MP |
| Flux 2 Create / Flux 2 Edit | The standard FLUX.2 endpoints in the same two modes | Cost-conscious creation and editing | see BFL pricing |
| Flux 2 Flex | FLUX.2 [flex], tunable steps/guidance | Typography and text rendering — the family's best | $0.05–0.06/MP |
| Flux 2 Max | FLUX.2 [max], the flagship | Grounded, fact-aware images; up to 10 references | $0.07/MP |
| Flux 2 Klein | FLUX.2 klein (4B and 9B) | Sub-second drafts, iteration, self-hosting | $0.014–0.015/image |
| Flux Kontext Pro | FLUX.1 Kontext [pro] | Fast conversational editing | $0.04/image |
| Flux Kontext Max | FLUX.1 Kontext [max] | Highest-fidelity Kontext edits, better typography | $0.08/image |
| Flux 2 LoRA Gallery | Community and custom LoRA adapters served on klein | Consistent characters, products, and house styles | varies |
| FLUX Pro Virtual Try-On | Dedicated garment try-on tool (v2) | E-commerce apparel imagery, up to 4 layered garments at 4MP | tool pricing |
| Control Light | Lighting-control tool in the Flux Tools line | Relighting a subject or scene without regenerating it | tool pricing |
The practical read: pro at $0.03/MP is the default, and the only reasons to pay up are flex's typography and max's grounded references. Prices are Black Forest Labs' official rates; third-party resellers quote conflicting numbers and are not a reliable source.
What are Flux 2's specs and limits?
| Spec | Value (as of Aug 2026) |
|---|---|
| Max resolution | 4MP (e.g. 2048×2048 and non-square equivalents) |
| Reference images | Up to 10 (max tier) |
| Color control | Exact hex-code binding (e.g. "background #0F62FE") |
| Structured prompting | JSON-formatted prompts supported |
| Negative prompts | Not supported — describe what you want, not what to avoid |
| Text encoder | Mistral-3 24B VLM (klein uses Qwen3-8B) |
| Architecture | Latent flow matching, single create+edit checkpoint |
The missing negative-prompt field trips up people arriving from older diffusion workflows: with Flux 2 you steer entirely through positive description, reference images, and hex/JSON constraints.
Who can actually use Flux 2 commercially? (The license split most articles get wrong)
Correction worth highlighting: among the open-weights Flux 2 checkpoints, only klein-4B ships under Apache 2.0 — the one genuinely unrestricted option. Both klein-9B and the 32B dev checkpoint use the FLUX Non-Commercial License: you may use their outputs commercially, but deploying or serving the model itself in a commercial product requires a paid license from Black Forest Labs. This split is widely misreported — many posts describe dev-32B as "open for commercial use," which conflates output rights with model-deployment rights.
Everything generated through the official API (pro, flex, max, hosted klein, Kontext) is commercially usable, per BFL's licensing terms as of August 2026. If you are self-hosting for a business, klein-4B is the only checkpoint that requires no paid agreement.
How good is Flux 2, honestly?
On the Artificial Analysis image arena (Elo, as of August 2026): Flux 2 max scores 1196, pro 1187, flex 1180, dev 1154, klein-9B 1120, and klein-4B 1057. That places the whole family behind the closed frontier — GPT Image 2 at 1339, Reve 2.1 at 1299, Nano Banana 2 around 1262.
So the honest case for Flux 2 is not "best image model." It is three narrower claims that hold up: dev-32B and the klein models are the strongest open-weights options in their size classes; the per-megapixel pricing ($0.03/MP at pro tier) delivers unusual price-performance; and it is the only family in this quality band you can legally fine-tune and self-host (within the license terms above). Enterprise adoption reflects that positioning — Flux 2 powers Envato's creative engine and is distributed through Cloudflare Workers AI and Azure AI Foundry.
What makes Flux 2 max different from every other image model?
Grounding web search. FLUX.2 [max] can query the web at generation time and condition the image on what it finds — current product designs, real logos and liveries, up-to-date visual facts. Combined with its 10-reference-image ceiling, this makes max the variant for images that must be correct, not merely attractive. As of August 2026 no other major image model ships retrieval-grounded generation as a first-party feature.
How do you prompt Flux 2?
Black Forest Labs' prompting guide recommends a Subject + Action + Style + Context structure at roughly 30–80 words, with the most important elements front-loaded. Practical rules:
- Bind colors by hex. "A running shoe on a plain #F4F1EA studio sweep, accent laces in #E4572E" reproduces brand palettes far more reliably than color names.
- Quote text exactly. For typography, put the exact string in quotes — 'a poster headline reading "SUMMER SALE — 40% OFF"' — and prefer the flex tier, the family's best text renderer.
- Use JSON for repeatable jobs. Structured prompts (subject, style, palette, layout as fields) make batch generation consistent.
- No negative prompts. Re-describe the scene positively instead of listing exclusions.
Example: "Product photo of a matte ceramic pour-over coffee set, steam rising, on a walnut counter with morning side-light, background #F7F3EC, minimalist Kinfolk editorial style, 4:5 crop."
What can you build with Flux 2?
The 2026 Tools line turns the family into a production stack for commerce and design work. Virtual Try-On v2 dresses a model image in up to four layered garments at 4MP, at catalog scale; Erase removes objects cleanly; Outpainting extends frames to 4MP for new aspect ratios. Typical workflows:
- Product and catalog imagery — try-on plus hex-locked backgrounds cover most apparel and DTC shot lists; see invideo's product photography workflow for how this runs inside the agent.
- Marketing and social visuals — pro-tier speed and cost make it a workhorse inside an AI image generator pipeline, from thumbnails to ad variants.
- Stylized and illustrative work — LoRA adapters on klein hold a consistent character or house style across a series, useful for AI art generation projects that need coherence, not one-offs.
Common questions about Flux 2
Is Flux 2 free for commercial use?
Outputs from the official API are commercially usable across all tiers. For self-hosted weights, only klein-4B (Apache 2.0) is unrestricted; klein-9B and dev-32B allow commercial outputs but require a paid BFL license to deploy the model commercially (as of August 2026).
Can you train a LoRA on Flux 2?
Yes. Black Forest Labs publishes LoRA training docs for klein: adapters run 10–200MB and train in roughly 1–3 hours on a consumer GPU using AI-Toolkit or Diffusers. Undistilled "klein Base" checkpoints exist specifically for fine-tuning, and you can upload trained LoRAs to the BFL API for hosted inference — no self-hosting needed.
How much does Flux 2 cost?
Official BFL pricing as of August 2026: pro $0.03/MP for creation and $0.045/MP for edits; flex $0.05–0.06/MP; max $0.07/MP; klein $0.014–0.015 per image; Kontext pro $0.04 and Kontext max $0.08 per image. Ignore reseller price lists — they contradict each other.
Does Flux 2 support negative prompts?
No. The family has no negative-prompt field. Steer with positive description, hex color codes, JSON-structured prompts, and reference images instead.
What's the difference between Flux 2 and Flux Kontext?
Kontext (May 2025) pioneered instruction-based editing as a separate model line. FLUX.2 (November 2025) folded creation and editing into one checkpoint. Kontext pro/max remain available and are still strong for fast conversational edits, but Flux 2's edit mode is the current-generation path.
How many reference images can Flux 2 use?
Up to 10 on the max tier — useful for locking identity, product geometry, and style simultaneously in one generation.
Where can you use Flux 2?
The entire family ships inside invideo, the AI video platform that gives serious creatives every major model in one place — the picker entries in the table above are live in the agent alongside 200+ other models, so you can generate a Flux 2 still and cut it straight into a video sequence without changing tools. Start at invideo.io, read the Flux Kontext model hub for the editing line, or jump into a workflow like the AI design generator or AI image generator.
Version history
- August 2026 — First published. Covers FLUX.2 pro/flex/max/dev/klein, Kontext pro/max, 2026 Tools (Virtual Try-On v2, Erase, Outpainting), Aug 2026 arena standings, and official pricing.