FLUX.2

FLUX.2 is Black Forest Labs's second-generation flagship text-to-image generation model family, developed by the original creators of Stable Diffusion and released in late 2025 following the FLUX.1 product line established at the company's founding in August 2024. The model generates high-resolution images from natural-language prompts, with capabilities including up-to-4K output resolution, multi-image reference synthesis, improved text rendering inside generated images, and higher photorealism across a range of scene categories. As of April 2026, FLUX.2 is widely characterized as one of the leading image-generation systems available to developers and enterprises, with a distinctive multi-tier release structure that spans a closed-weights commercial flagship and open-weights research and inference variants.

At a glance

• Lab: Black Forest Labs
• Released: Late 2025
• Modality: Image (text-to-image generation)
• Open weights: Mixed. FLUX.2 Pro is closed-weights, gated through the BFL commercial API. FLUX.2 Dev is released as open weights under a non-commercial research license. FLUX.2 Schnell is released under the Apache 2.0 license, permitting commercial use without license fees.
• Output resolution: Up to 4K pixels (flagship Pro variant); standard-resolution targets available for Dev and Schnell variants
• Pricing: Schnell: free for self-hosted use (Apache 2.0). Dev: non-commercial open weights; commercial deployment requires a separate license agreement. Pro: per-image pricing through the BFL API at https://api.bfl.ai and partner platforms; see current rate card at api.bfl.ai.
• Distribution channels: Hugging Face black-forest-labs organization (https://huggingface.co/black-forest-labs) for Dev and Schnell; BFL commercial API (https://api.bfl.ai) for Pro; partner integrations including Together AI, Replicate, and fal.ai; self-hosting via Diffusers and ComfyUI for open-weights variants

Origins

Black Forest Labs was founded in August 2024 by Robin Rombach, Andreas Blattmann, Patrick Esser, and Dominik Lorenz, the principal researchers behind Stable Diffusion at the University of Munich's Computer Vision Group (CompVis) and at Stability AI. Rombach, Blattmann, and Esser were lead authors of "High-Resolution Image Synthesis with Latent Diffusion Models" (December 2021), the research paper that introduced the latent diffusion architecture underlying Stable Diffusion, and they continued developing Stable Diffusion variants at Stability AI through 2024.

The founding of Black Forest Labs followed a period of organizational difficulty at Stability AI through 2023 and 2024. The core Stable Diffusion team departed Stability to establish an independent research lab, with a $31 million Series A led by Andreessen Horowitz announced simultaneously with the company's launch. The first product release, FLUX.1, arrived in August 2024 concurrent with the company's public debut.

FLUX.1 established the three-variant release pattern. FLUX.1 Pro, the closed-weights commercial flagship, was distributed through the BFL API and partner platforms. FLUX.1 Dev, released as open weights under a non-commercial license through Hugging Face, provided research and developer access to the model weights. FLUX.1 Schnell, also open weights, was released under the Apache 2.0 license, permitting commercial use without fees, with a faster and lighter architecture optimized for inference throughput. Industry coverage at launch characterized FLUX.1 as superior to contemporary Stable Diffusion variants and competitive with or superior to Midjourney V6 and DALL-E 3 on most image-quality evaluation categories.

FLUX.1.1 Pro arrived in October 2024 with improvements to image quality and generation speed. FLUX.1 Kontext, released in 2025, extended the family to image editing through text-prompt-driven modification of existing images.

FLUX.2 launched in late 2025 as the second-generation flagship. The release introduced improvements to text rendering, photorealistic rendering quality, multi-image reference synthesis (using up to 10 input images for style and tone consistency), and output resolution capability up to 4K pixels. FLUX.2 retained the same three-variant pattern as FLUX.1: Pro (closed-weights), Dev (open-weights non-commercial), and Schnell (Apache 2.0). The December 2025 Series B of $300 million at a $3.25 billion valuation, co-led by Salesforce Ventures and Anjney Midha (AMP), reflected commercial traction from the FLUX family.

Capabilities

FLUX.2's architecture is based on a rectified-flow latent diffusion approach. Rectified flow training arranges the trajectory from noise to image in a more linear path than earlier diffusion schedules, improving training efficiency and enhancing generated image quality at equivalent compute budgets. The latent diffusion framing compresses images into a lower-dimensional latent space for generation before decoding to full resolution, keeping compute tractable at high output resolutions.

The model's documented strengths fall into four categories. Prompt fidelity: FLUX.2 reliably represents multi-element text descriptions, including correct counts of objects, specified spatial arrangements, style attributes, and multiple interacting subjects. Photorealism: FLUX.2 produces detailed, naturalistic images with accurate lighting, materials, and textures across a wide range of scene categories. Text rendering inside generated images: FLUX.2 improves on FLUX.1's performance on rendering legible text within scenes, relevant for marketing materials, signage, typographic design, and other content where text appears within the image. Multi-image reference: FLUX.2 Pro accepts up to 10 reference images as conditioning inputs, allowing style, tone, and visual identity to be specified through example rather than text description alone.

The three variants offer different trade-offs. FLUX.2 Pro, the closed-weights commercial flagship, prioritizes maximum output quality and is the primary target for the highest-resolution and most photorealistic generation tasks. FLUX.2 Dev, the open-weights research variant, provides model weights for non-commercial research, fine-tuning experiments, and developer exploration without license fees, at a quality level close to Pro. FLUX.2 Schnell, the Apache-2.0-licensed inference-optimized variant, trades some quality for substantially faster generation speed, suitable for high-throughput or cost-sensitive applications where per-image generation speed is a significant constraint.

Benchmarks and standing

Image-generation benchmarking is substantially less standardized than text-model benchmarking. There is no widely adopted composite leaderboard equivalent to the Artificial Analysis Intelligence Index. Evaluations typically combine human-preference side-by-side comparisons, FID (Frechet Inception Distance) scores measuring distributional similarity, and capability-specific tests covering text rendering accuracy, prompt adherence, and photorealism ratings across prompt categories.

The Hugging Face Text-to-Image Leaderboard (https://huggingface.co/spaces/google/image-gen-eval) evaluates models on prompt adherence and image quality across standardized prompt sets. FLUX.1 Pro held a top position on the leaderboard through much of late 2024 and into 2025 among both open-weights and closed commercial systems. FLUX.2 maintains this competitive position into 2026.

LMArena operates an image arena alongside its text leaderboards, with human raters evaluating pairwise image comparisons. FLUX.2's standing in the image arena positions it among the top tier of available systems, alongside Midjourney V7 and Imagen 4 Ultra. DALL-E 3 generally trails on most quality dimensions in the arena evaluations.

Among the leading image-generation systems as of April 2026, the competitive field at the top of quality rankings includes Midjourney V7 (dominant on stylized and artistic outputs), Imagen 4 Ultra (strong on photorealism and text rendering in the enterprise Google Cloud context), and FLUX.2 (leading on most composite quality categories, with particular strength in prompt fidelity and photorealism, plus the open-weights advantage). DALL-E 3 retains the largest user base through ChatGPT distribution but trails on quality dimensions in head-to-head comparisons.

FLUX.2's distinctive position in the benchmark landscape is as the highest-quality model available in open-weights form, bringing frontier image-generation quality within reach of researchers and developers without requiring a commercial API relationship.

Benchmark leadership in image generation is point-in-time: methodologies are not standardized, and a new model release can reshape the leaderboard within weeks.

Access and pricing

Access to FLUX.2 varies by variant.

FLUX.2 Schnell and FLUX.2 Dev are available through the black-forest-labs organization on Hugging Face (https://huggingface.co/black-forest-labs). Schnell is released under the Apache 2.0 license and can be downloaded, self-hosted, and used commercially without license fees. Dev is released under a non-commercial license; commercial deployment requires a separate license agreement with Black Forest Labs. Both variants load with the Diffusers library and run in ComfyUI with the relevant checkpoint loader.

FLUX.2 Pro is available through the BFL commercial API at https://api.bfl.ai, with per-image pricing published there and subject to revision. Partner platforms including Together AI (https://www.together.ai), Replicate (https://replicate.com), and fal.ai (https://fal.ai) integrate FLUX.2 Pro through their own API surfaces and pricing models. Adobe Firefly integration brings FLUX-derived capabilities into Adobe Creative Cloud for qualifying customers.

Comparison

The direct peer set for FLUX.2 in April 2026 is the leading text-to-image generation systems:

• DALL-E 3 (OpenAI). The image-generation model with the largest consumer user base, distributed through ChatGPT and Microsoft Bing Image Creator. DALL-E 3 leads on distribution breadth; FLUX.2 leads on most quality dimensions. DALL-E 3's ChatGPT integration gives it structural reach among non-specialist users who are already in that interface; FLUX.2 requires API access or a partner integration for Pro, though the open-weights Schnell variant is accessible to any developer.
• Imagen 4 (Google DeepMind). Google's fourth-generation image-generation model, available through Vertex AI, the Gemini app, and Google AI Studio. Imagen 4 Ultra is competitive with FLUX.2 Pro on photorealism and text rendering, and holds a structural advantage through Google Cloud's enterprise distribution infrastructure and SLA commitments. FLUX.2's advantages include the open-weights variants, broader partner integration across non-Google platforms, and the multi-image reference capability.
• Midjourney V7. The dominant prosumer image-generation system with a subscriber base generating approximately $500 million in annual revenue. Midjourney V7 leads on stylized, artistic, and aesthetic-quality outputs where a distinctive look is the goal. FLUX.2 leads on prompt fidelity for complex multi-element descriptions and on photorealistic scene generation. Midjourney distributes through Discord and its own web product; it offers no open-weights access.
• Stable Diffusion 3.5 (Stability AI). Stability AI's current open-weights flagship, available on Hugging Face and through the Stability AI API. The relationship between Black Forest Labs and Stability AI is that of successor and prior employer: the FLUX.2 team built Stable Diffusion, then left to create what they consider the next generation of the open-weights image generation lineage. Stable Diffusion 3.5 generally trails FLUX.2 on composite quality benchmarks and community adoption metrics.

FLUX.2's distinctive position across this peer set is the combination of open-weights frontier quality and a three-tier license structure bridging research, non-commercial, and commercial use cases. No competing system offers both the quality ceiling of FLUX.2 Pro and open-weights access at the Dev and Schnell tiers.

Outlook

Several open questions shape FLUX.2's trajectory through 2026 and into 2027:

• FLUX.3 timing and capability profile. Black Forest Labs has maintained a rapid release cadence since the company's founding: FLUX.1 in August 2024, FLUX.1.1 Pro in October 2024, FLUX.1 Kontext in 2025, and FLUX.2 in late 2025. A FLUX.3 or next-generation release in 2026 is plausible given this pace. The capability questions for a successor are whether image editing integration deepens, whether the 4K resolution ceiling extends further, and whether the open-weights quality gap to Pro narrows.
• BFL API monetization and commercial scale. FLUX.2 Pro's commercial API is the primary revenue vehicle for Black Forest Labs alongside partner integration fees. The December 2025 Series B at $3.25 billion was predicated on commercial traction from this model. Whether the API achieves the scale of usage that justifies the valuation will depend on developer and enterprise adoption in 2026.
• The open-weights versus proprietary capability gap. A structural question in the image-generation market is whether open-weights models can maintain parity with closed proprietary systems as training compute and data investments scale. FLUX.2's position, where the open-weights Dev variant approaches but trails Pro on quality, represents the current answer. Whether future FLUX releases narrow that gap further, or whether the proprietary frontier pulls away, is an open question.
• FLUX.2 Pro's competitive position against closed peers. Midjourney, OpenAI, and Google DeepMind are all continuing to advance their image-generation capabilities. The closed-weights Pro tier's quality advantage over these peers is not structurally guaranteed; maintaining it requires continued training investment and architectural advancement.
• Nemotron Coalition collaborative output. Black Forest Labs joined the Nemotron Coalition convened by NVIDIA Research in March 2026 as one of eight inaugural members. How the collaboration's research feeds into future FLUX releases is not yet established.

Sources

• Black Forest Labs official site. Company and product overview.
• BFL commercial API. FLUX.2 Pro access, rate card, and developer documentation.
• Black Forest Labs on Hugging Face. Open-weights FLUX.2 Dev and Schnell model distribution.
• Wikipedia: Flux (text-to-image model). Model and company history, FLUX lineage from Stable Diffusion.
• VentureBeat: Stable Diffusion creators launch Black Forest Labs. Founding context, Series A, FLUX.1 launch coverage.
• TechCrunch: Black Forest Labs raises $300M at $3.25B valuation. Series B coverage and commercial trajectory.
• Hugging Face Text-to-Image Leaderboard. Cross-model image-generation benchmark comparison.