FLUX.2: 32B Image Model, 4MP Output, $0.03/Image (2026)
FLUX.2 by Black Forest Labs (Nov 2025): 32B model, 4MP output, $0.03/image. Leads 2026 photorealism with multi-reference conditioning and text rendering.
FLUX.2 is Black Forest Labs' second-generation image model (November 25, 2025), built on a 32B rectified flow transformer that generates images up to 4 megapixels with multi-reference conditioning across up to 10 reference images and sub-10-second generation time. FLUX.2 [pro] costs $0.03 for the first megapixel plus $0.015 per additional MP, with the open-weight FLUX.2 [dev] 32B checkpoint available non-commercially on Hugging Face.
FLUX.2, released November 25, 2025 by Black Forest Labs, is a 32B rectified flow transformer for text-to-image generation producing images up to 4 megapixels. It supports multi-reference conditioning across up to 10 reference images, exact hex color matching, and sub-10-second generation. FLUX.2 [pro] costs $0.03 for the first megapixel plus $0.015 per additional MP; the FLUX.2 [dev] 32B model is available open-weight on Hugging Face for non-commercial use.
Provider: Black Forest Labs · Family: FLUX.2
Input modalities: text, image · Output: image
About FLUX.2
FLUX.2 is a family of text-to-image generation models built by Black Forest Labs, the German-American AI research company co-founded in 2024 by Robin Rombach, Andreas Blattmann, Patrick Esser, and Dominik Lorenz. All four co-founders previously built Stable Diffusion at Stability AI before departing in early 2024. FLUX.2 launched on November 25, 2025, as the second-generation successor to the widely adopted FLUX.1 series. The flagship variants are built on a rectified flow transformer architecture with 32 billion parameters, making FLUX.2 Dev and Pro one of the largest publicly accessible image generation models available. The family spans five distinct variants: FLUX.2 [max] for cinematic photorealism, FLUX.2 [pro] for production API workloads, FLUX.2 [flex] for developer-tunable inference, FLUX.2 [dev] as a 32B open-weight checkpoint, and FLUX.2 [klein] in compact 4B and 9B sizes for sub-second consumer-GPU generation. FLUX.2 is positioned as the 2026 photorealism leader among commercial image generation models, outperforming FLUX.1 and Midjourney v6 on photorealism and prompt adherence in third-party comparisons. Multiple independent reviewers tested FLUX.2 Pro against Midjourney v7, GPT Image 2, and Google Imagen 4 across product photography, portrait generation, and typographic use cases. FLUX.2 Pro ranks first or second across 2026 photorealism leaderboards; Midjourney v7 retains the lead on artistic stylization; Imagen 4 edges out FLUX.2 on complex text rendering within images. FLUX.2 is 30 to 50 percent faster at image generation than FLUX.1, bringing Pro API generation times under 10 seconds for standard 1MP outputs. No single standardized numeric benchmark exists for commercial image generation the way SWE-bench covers code, but FLUX.2 Pro leads category-level photorealism rankings across at least five independent 2026 reviewer sites. FLUX.2 generates images at native resolutions up to 4 megapixels, a doubling of the 2MP ceiling in FLUX.1. All Pro, Max, and Flex variants support a wide range of aspect ratios. FLUX.2 Pro handles up to 10 reference images per generation call, maintaining character, style, and environment consistency across the full set. FLUX.2 [dev] supports 6 reference images; FLUX.2 [klein] supports up to 4. A redesigned Variational Autoencoder in FLUX.2 addresses geometry and texture degradation that caused quality loss in FLUX.1 at high resolutions. Klein (4B) fits in 13GB VRAM, making sub-second generation possible on consumer RTX 3090 and RTX 4070 cards. FLUX.2 accepts text prompts and image references as inputs, and outputs static images only. Multi-reference conditioning lets developers specify a subject from one image, a style from a second, and an environment from a third in a single generation call. All variants support exact hex color matching, enabling precise brand color reproduction without post-processing correction. JSON-structured prompting is supported alongside natural language, giving precise control over scene elements, camera settings, lens type, and composition for repeatable commercial outputs. FLUX.2 Pro and Max support pose guidance for consistent character positioning. FLUX.2 [max] adds grounded generation, which incorporates real-time web context into generated images. FLUX.2 does not generate audio, video, or text output. Image editing (not just generation) is supported by FLUX.2 [dev] and FLUX.2 [klein]. Black Forest Labs uses credit-based pricing where 1 credit equals $0.01 USD, with megapixel-based pricing scaling cost to output resolution. FLUX.2 [pro] charges $0.03 for the first megapixel and $0.015 for each additional megapixel, so a full 4MP image costs $0.075. FLUX.2 [flex] is the most affordable API variant at roughly $0.01 per 1MP image. FLUX.2 [dev] costs approximately $0.012 per image through the API, with open weights also on Hugging Face for non-commercial self-hosting. FLUX.2 [klein] is priced at about $0.014 per image through the API. FLUX.2 [max], the highest-quality tier, is priced above Pro with costs calculated via the pricing calculator at bfl.ai/pricing. For a team generating 10,000 product images per month at 1MP using FLUX.2 [pro], the cost is $300 per month. The same volume through FLUX.2 [flex] costs approximately $100 per month. FLUX.2 is accessible through the Black Forest Labs API at api.bfl.ai, with a Python client library and standard REST endpoints. Third-party inference providers including fal.ai, Replicate, DeepInfra, and CometAPI host FLUX.2 variants, often at rates competitive with or below the direct BFL API. FLUX.2 [dev] (32B) and FLUX.2 [klein] (4B and 9B) weights are on Hugging Face at huggingface.co/black-forest-labs under BFL's non-commercial open weights license, requiring acceptance of model card terms before downloading. Full-precision 32B dev weights require an H100 or equivalent GPU (80GB VRAM); quantized FP8 versions run on an RTX 4090 (24GB VRAM). FLUX.2 [klein] 4B fits in 13GB VRAM on RTX 3090 and above. The BFL playground at playground.bfl.ai provides no-code browser access for testing all variants. Black Forest Labs published a usage policy and model card for FLUX.2 prohibiting generation of CSAM, non-consensual intimate imagery, and deceptive synthetic media of real people without consent. The API applies automated content filters enforcing these restrictions, and the open weights license binds self-hosting users to the same prohibitions. BFL has not published a formal red-team evaluation with named external partners as of June 2026, which distinguishes it from frontier language model providers such as Anthropic and Google who detail independent safety evaluations in published system cards. Safety filtering on the API is configurable by enterprise customers. FLUX.2 Pro is the right choice for e-commerce product photography, marketing creative, and UI mockup generation where photorealism and brand color accuracy matter more than artistic flair. FLUX.2 Flex suits developer teams who need direct control over inference steps and guidance scale for domain-specific quality targets. FLUX.2 Klein is best for consumer apps requiring real-time generation on modest hardware. Teams building highly artistic or stylized campaign imagery should evaluate Midjourney v7, which still leads in non-photorealistic artistic output. Teams requiring native video or audio generation must use other models entirely, as FLUX.2 outputs static images only. Developers needing enterprise compliance certifications (SOC 2 Type II, HIPAA) should contact BFL sales directly, as the standard API does not publish these certifications as of June 2026. Black Forest Labs has not published a detailed training data card for FLUX.2 specifying exact dataset composition or a precise training cutoff date. The model was trained by the BFL team on compute clusters across Europe and the United States. API data retention and training practices for API inputs are covered in BFL's terms of service at bfl.ai/legal/terms. The FLUX.2 Dev open weights license explicitly restricts commercial SaaS deployment without a separate commercial license above 100,000 images per month. Enterprises requiring higher volume from open weights must negotiate terms with BFL. The company is headquartered in Germany and operates under EU AI Act obligations as a general-purpose AI system provider. FLUX.2 is a significant upgrade over FLUX.1, which launched in August 2024. Key improvements: native 4MP resolution (up from 2MP), a redesigned VAE for cleaner geometry at high resolution, multi-reference conditioning across up to 10 images, 30 to 50 percent faster generation, exact hex color matching, JSON-structured prompting, and improved text rendering for UI mockups and infographics. Black Forest Labs raised its $300 million Series B concurrent with the FLUX.2 launch in December 2025 at a $3.25 billion valuation. The Klein variants (4B and 9B) were new in FLUX.2, introduced to serve consumer-facing applications needing sub-second generation on RTX 3090-class hardware. No FLUX.3 roadmap has been announced as of June 2026, but BFL has signaled continued investment in higher resolution, video generation, and real-time generation capabilities.
Pricing
Credit-based pricing: 1 credit = $0.01 USD. Megapixel-based: FLUX.2 [pro] $0.03 first MP + $0.015 per additional MP (full 4MP = $0.075). FLUX.2 [flex] ~$0.01/image. FLUX.2 [dev] ~$0.012/image. FLUX.2 [klein] ~$0.014/image. FLUX.2 [max] priced above Pro; see bfl.ai/pricing for calculator.
Key Features
- Multi-Reference Conditioning: Combines up to 10 reference images (Pro/Max) in a single call, using weighted attention to independently transfer subject identity, style, and environment without cross-contamination between references.
- 4MP Native Resolution: Generates images at up to 4 megapixels natively, doubling FLUX.1's 2MP ceiling. A redesigned VAE ensures clean geometry and texture at high resolution without the degradation seen in first-generation models.
- Exact Hex Color Matching: Accepts specific hex color codes in prompts and reliably reproduces them in outputs, enabling brand-accurate product visuals without manual color correction post-processing.
- JSON-Structured Prompting: Accepts structured JSON alongside natural language prompts, allowing precise specification of camera settings, scene elements, lighting, and composition for consistent commercial image pipelines.
- FLUX.2 [klein] Sub-Second Generation: The 4B and 9B Klein variants generate images in under 1 second on RTX 3090/4070+ (13GB VRAM), enabling real-time image generation for consumer-facing applications at $0.014/image.
- Open-Weight Dev Model: FLUX.2 [dev] at 32B parameters is on Hugging Face for non-commercial use, allowing researchers and fine-tuners to experiment with and adapt the full frontier model without API costs.
Pros
- Leads 2026 photorealism comparisons against Midjourney v7, Imagen 4, and GPT Image 2 on product photography and anatomy rendering in multiple independent reviews.
- Multi-reference conditioning (up to 10 refs for Pro/Max) maintains character, style, and environment consistency in a single generation call, a capability absent from most competitors at launch.
- Five-variant family spans $0.01/image consumer to $0.075/4MP enterprise, with open-weight Klein and Dev enabling self-hosted deployment on consumer RTX hardware from 13GB VRAM.
Cons
- Static image output only: no audio, video, or text generation — voice and video workflows require integrating separate model providers.
- No independent red-team safety report published as of June 2026; safety posture is self-attested via model card, unlike Anthropic and Google who publish named third-party audit results.
- FLUX.2 Dev open weights carry a non-commercial license; SaaS deployment above 100K images/month requires a paid BFL commercial license, catching teams who prototype with open weights and scale to production.
Benchmarks
Frequently Asked Questions
What is FLUX.2 and who built it?
FLUX.2 is a family of text-to-image generation models released November 25, 2025 by Black Forest Labs, a German-American AI research company co-founded by Robin Rombach, Andreas Blattmann, Patrick Esser, and Dominik Lorenz. All four co-founders previously built Stable Diffusion at Stability AI before departing in early 2024. FLUX.2 is built on a rectified flow transformer architecture, a departure from traditional DDPM diffusion that produces cleaner geometry with fewer denoising steps. The flagship variants carry 32 billion parameters, and a compact Klein tier offers 4B and 9B models for consumer hardware. FLUX.2 includes five variants: Pro (production API), Max (cinematic quality), Flex (developer-tunable), Dev (32B open weights), and Klein (sub-second, compact). The model family succeeds FLUX.1, which launched in August 2024 and became the most downloaded image generation model family on Hugging Face. FLUX.2 was released concurrent with Black Forest Labs' $300M Series B at a $3.25B valuation in December 2025.
How much does FLUX.2 cost per image?
FLUX.2 uses megapixel-based credit pricing where 1 credit = $0.01 USD. FLUX.2 [pro] charges $0.03 for the first megapixel plus $0.015 for each additional megapixel, so a full 4MP image costs $0.075. FLUX.2 [flex] costs approximately $0.01 per 1MP image, making it the most affordable API tier. FLUX.2 [dev] costs around $0.012 per image via the BFL API, and the 32B weights are also free on Hugging Face for non-commercial use. FLUX.2 [klein] (4B/9B) is priced at approximately $0.014 per image through the API. FLUX.2 [max] is priced above Pro; use the bfl.ai/pricing calculator for exact costs. For scale: 10,000 product images at 1MP via FLUX.2 [pro] costs $300/month; the same volume through FLUX.2 [flex] costs roughly $100. Third-party providers (fal.ai, Replicate, DeepInfra) often match or beat BFL direct pricing. There is no free tier; all API usage is credit-billed.
What resolution and output does FLUX.2 support?
FLUX.2 generates static images at native resolutions up to 4 megapixels (4MP), doubling the 2MP ceiling of FLUX.1. FLUX.2 supports a wide range of aspect ratios; the Flex variant exposes aspect ratio controls directly to developers. The model cannot generate video, audio, or text output. Image editing (not just generation) is supported by FLUX.2 [dev] and FLUX.2 [klein]. A redesigned Variational Autoencoder in FLUX.2 addresses the geometry and texture degradation that appeared in FLUX.1 at high resolutions, ensuring clean surfaces and accurate anatomy at 4MP. Klein variants (4B/9B) also support up to their reference image count limits at full 4MP resolution. There is no sliding context window concept for image models: each generation call is independent, with reference images provided per call rather than accumulated across a session.
How does FLUX.2 compare to Midjourney and other image AI models in 2026?
Multiple independent 2026 reviewers rank FLUX.2 Pro as the leader in photorealism and product photography, with Midjourney v7 still winning on artistic stylization and non-photorealistic output. Google Imagen 4 (released April 2026) edges out FLUX.2 specifically on complex text rendering within images, such as precise multi-word typography. GPT Image 2 from OpenAI combines image quality with ease of use but lags behind FLUX.2 Pro on raw photorealism for commercial product visuals. FLUX.2 distinguishes itself with multi-reference conditioning (up to 10 reference images in Pro) and exact hex color matching, which neither Midjourney nor Imagen 4 supports natively at the same fidelity. FLUX.2 is 30 to 50 percent faster than FLUX.1 but is not the fastest overall: Klein variants achieve sub-second generation, while Pro takes 7 to 10 seconds. Teams should choose FLUX.2 Pro for photorealistic commercial imagery and Midjourney v7 for artistic, painterly, or cinematic creative direction.
Is FLUX.2 open source?
FLUX.2 uses a mixed openness model, not fully open source. FLUX.2 [dev] (32B parameters) and FLUX.2 [klein] (4B and 9B) are available on Hugging Face at huggingface.co/black-forest-labs under BFL's non-commercial open weights license. Research, personal projects, and non-commercial applications can use these weights freely after accepting the license terms on the model card. Commercial SaaS products that generate images for external users must negotiate a paid commercial license with Black Forest Labs if generating above 100,000 images per month from the open weights. FLUX.2 [pro], [max], and [flex] are proprietary API-only variants with no public weights released. The earlier FLUX.1 [schnell] was released under Apache 2.0 (fully open source) but FLUX.2 variants all use BFL's custom non-commercial license. Weight formats include safetensors with FP16 and FP8 quantization options.
What modalities does FLUX.2 support?
FLUX.2 accepts two input modalities: text prompts and image references. Output is static images only, at resolutions up to 4MP. FLUX.2 does not generate audio, video, or text. Multi-reference conditioning accepts up to 10 reference images per call (Pro/Max), 6 for Dev, and 4 for Klein, maintaining character, style, and environment consistency across the set. Function calling, structured JSON output for text, and tool use are not supported. JSON-structured prompting is supported as an input format for controlling scene elements and camera parameters, but this is a prompting interface, not an agent tool-use capability. Pose guidance is available for Pro and Max variants. Grounded generation, which incorporates real-time web context, is exclusive to FLUX.2 [max]. Image editing (modifying an existing image) is supported by Dev and Klein variants in addition to generation. FLUX.2 cannot browse the web independently except through the Max grounded generation feature.
Does FLUX.2 train on user data submitted via the API?
Black Forest Labs states it does not train on API inputs submitted by users. The company's terms of service at bfl.ai/legal/terms govern data handling for API customers, but specific retention period details are not publicly disclosed as of June 2026. There is no published zero-retention enterprise option equivalent to Anthropic's or OpenAI's enterprise offerings. BFL does not hold publicly disclosed SOC 2 Type II, HIPAA-eligible, or ISO 27001 certifications as of June 2026, which limits FLUX.2's suitability for regulated healthcare, legal, or government workloads without direct enterprise agreements. The company is headquartered in Germany and operates under GDPR. FLUX.2 [dev] open weights self-hosters control their own data entirely, with no API calls to BFL infrastructure, providing full data sovereignty for non-commercial research workflows.
Who should use FLUX.2 and who should avoid it?
FLUX.2 Pro is the best choice for e-commerce product photography teams, marketing creative agencies needing multi-reference brand consistency, and developers building image pipelines requiring precise hex color accuracy and high photorealism. FLUX.2 Flex suits ML engineers who need direct control over inference steps and guidance scale for domain-specific tuning. FLUX.2 Klein (4B/9B) is ideal for consumer-facing apps requiring sub-second image generation on RTX 3090+ hardware at 13GB VRAM. Teams building artistic or painterly campaign imagery should evaluate Midjourney v7 instead, which retains the quality lead on non-photorealistic styles. Teams needing audio or video generation should use purpose-built models (Runway, Kling, ElevenLabs). Enterprise teams in regulated industries requiring published SOC 2 or HIPAA compliance should contact BFL sales before committing, as these certifications are not publicly disclosed. Developers who want to fine-tune a frontier model cheaply should consider FLUX.2 Dev open weights, but must negotiate a commercial license before serving external users at scale.