FLUX.2
Black Forest Labs 推出的下一代图像生成与编辑模型家族，支持最多 10 张参考图的身份一致性融合和 4MP 原生分辨率输出
In-depth Report
-
FLUX.2, released by Black Forest Labs in November 2025, is one of the most comprehensive image generation and editing models currently available. It surpasses Midjourney V7 in photorealism, supporting identity-consistent fusion of up to 10 reference images, 4MP native resolution output, readable complex typography rendering, and ultra-long prompt word context of 32K tokens. At the same time, Black Forest Labs continued the core concept of open source and released the 32B parameter open weight version FLUX.2 [dev] and the lightweight version FLUX.2 [klein] of the Apache 2.0 protocol, making it the most active image generation base in the open source community after Stability AI.
-
Black Forest Labs was co-founded in 2024 by four former Stability AI core researchers, Robin Rombach, Andreas Blattmann, Patrick Esser and Dominik Lorenz - the creators of the latent diffusion model (Latent Diffusion) and the original architecture of Stable Diffusion. The company is headquartered in Freiburg, Germany (Black Forest region) and the San Francisco Bay Area, USA, and operates dual headquarters. Less than a year and a half after its establishment, BFL completed an amazing financing pace: in August 2024, it received a US$31 million seed round led by Andreessen Horowitz, General Catalyst and Y Combinator; then Series A was led by a16z, with participation from NVIDIA, Creandum, Earlybird and Northzone; in December 2025, it completed a US$300 million Series B, with a valuation of 32.5 US$ 430 million, with cumulative financing exceeding US$ 430 million. This makes BFL the best-funded pure-play image-generative modeling company in the world. The company's core business model is "Open Core" - the open weight of the front-end model greatly reduces the threshold for community use, and the back-end serves corporate customers through high-performance APIs to monetize. Leading technology companies such as Adobe, Meta and Canva are already its API and model licensing customers.
-
FLUX.2 is not a single model, but a family of models with multiple variants, covering the full spectrum from real-time inference to the highest quality: **Multi-reference image fusion** is the most prominent feature of FLUX.2. Users can upload up to 10 reference images at the same time, and the model can extract character, product, style, and composition features across images and synthesize them into an output that is highly consistent and non-compliance. In the past, this required repeated retouching or the use of multiple tools to achieve - for example, integrating the clothing and poses of five different characters into one product poster, FLUX.2 can complete it in one generation. **Photorealism** is another core advantage. In multiple blind tests, FLUX.2 [pro] outperformed Midjourney V7 in photorealism, especially in tasks that require realistic materials and lighting logic, such as portraits, product photography, and indoor scenes. FLUX.2 supports native resolution output up to 4MP for print quality without the need for additional upscaling. **Text rendering capabilities** are significantly improved compared to FLUX.1. Both Chinese and English text in complex typography, infographics, and UI prototypes remain readable. Post-processing is still recommended for small font sizes, but for needs such as the production of posters, investment pages, and information graphics, it has entered the stage of implementation. **32K token prompt word context** is a key design decision at the architectural level of this model. Users can use structured JSON input to detail the hierarchical composition of the foreground, middle ground, and background, specifying precise color numbers (Hex Code), lighting angles, and material characteristics. This is especially valuable for automated pipelines and large-scale batch production scenarios. **Precise Color Control** supports direct input of hexadecimal brand colors, and the model is reproduced with extremely high color accuracy. E-commerce brands can accurately control the main colors of product images and brand logo colors without the need for post-stage color correction.**Four levels of model selection**: - FLUX.2 [max]: highest quality, supports real-time network search based ground generation, suitable for final asset output - FLUX.2 [pro]: Best value for money production grade option, ~$0.03/1024x1024 image, 6-9 seconds to generate - FLUX.2 [flex]: For developers who need parameter control, the number of sampling steps and guidance ratio can be adjusted, about $0.06/MP - FLUX.2 [dev]: 32B open weight, can be deployed on your own GPU, free for non-commercial use, commercial license $999/month - FLUX.2 [klein]: Apache 2.0 protocol open source model, 4B/9B sizes, can run in real time on consumer GPU (about 13GB VRAM)
-
BFL adopts a model that combines pay-per-volume API pricing with open weights. The generation cost of FLUX.2 [pro] on the API side is about $0.03/image (1024x1024), which is much lower than Midjourney’s monthly subscription. FLUX.2 [klein] Starting at only $0.014/image on API. For self-hosted users, the open-weight version can be run locally at zero cost (just bring your own GPU computing power). BFL's monetization path is clear: seize the community's mind and developer ecology through open weight, and obtain corporate payments through high-performance APIs. Its enterprise licensing customers include Adobe, Meta and Canva, proving that the model is being recognized at the high end of the market.
-
Positive reviews focused on photorealism and multi-reference fusion capabilities. The professional review website ThePlanetTools.ai gave an overall score of 9.2/10, saying that FLUX.2 [pro] "outperformed Midjourney V7 in photorealism" in a blind test. Chinese media and communities also said that FLUX.2’s progress in multi-reference consistency and readable text rendering is “obviously perceptible” and is closer to the “implementable” state than the FLUX.1 generation. The developer community has responded enthusiastically to the open weight version, and FLUX.2 has gradually replaced Stable Diffusion as the default base in the ComfyUI ecosystem. Negative feedback is mainly concentrated in several aspects: FLUX.2 [pro] is an API-specific model and has no open weight; the self-hosted [dev] version requires higher computing power (RTX 5090 or H100 level GPU is recommended); compared to Midjourney's artistic stylized output, FLUX.2 is more biased towards photographic realism and is slightly inferior in painting and stylization. In addition, the size of the community is still growing rapidly, and there is still a gap in the number of plug-ins and auxiliary tools in the Stable Diffusion ecosystem.
-
The industry generally believes that FLUX.2 is the most influential image model release in the first half of 2026. BFL's open source strategy is seen as the inheritance and upgrade of the Stability AI roadmap - it is no longer just "release the model and run", but is equipped with a complete API, documentation, enterprise support and developer tool chain. In terms of product positioning, FLUX.2’s direct competitors include Midjourney V7 (leading in artistic style), Ideogram 3 (leading in small font typesetting), Adobe Firefly (deep integration with design tools) and Google Imagen 3 (ecosystem binding). The core differentiation of FLUX.2 is: open innovation based on open source + photorealism + multi-reference fusion capability + structured prompt word support. It is neither a purely open source community project nor a purely closed source commercial API, but a combination of both.
-
The open source licensing aspect deserves attention. FLUX.2 [dev] adopts a non-commercial open-weight license, and enterprise commercial use requires a monthly license fee of $999. This threshold is not low for small and medium-sized startups, and there are also some criticisms in the community that the actual degree of "open source" is not as good as expected. Another potential risk is ethics and compliance issues. Photorealism makes AI-generated content more difficult to identify and may be abused for deepfakes, false advertising, and fraud scenarios. Although BFL has issued a content safety statement, it lacks a strict content moderation mechanism like Midjourney or Adobe Firefly. In addition, BFL faces extremely competitive market conditions. Midjourney is working on its V8 version, Stability AI has a new open source model roadmap, and both Google and Meta are increasing their investment in image generation. BFL has a first-mover advantage when financing is abundant, but the long-term competitive landscape remains unclear.
-
FLUX.2 is most suitable for the following groups: e-commerce and brand marketing teams (batch generation of product images, precise control of brand colors), advertising creative agencies (multi-reference fusion, role consistency), UI/UX designers (interface prototypes and infographic production), and enterprises that require local deployment (self-hosting using [dev] or [klein] versions). It is not suitable for the following people: creators who pursue stylized output of pure art (Midjourney is recommended), workflows that rely on the community plug-in ecosystem (Stable Diffusion is still recommended), and individual users with limited computing power ([pro] API costs are controllable, but the threshold for local deployment is high). Recommendations for use: Choose FLUX.2 [pro] for most production scenarios, which is the most cost-effective; upgrade to [flex] for scenarios that require precise typesetting; choose [dev] for local deployment; choose [klein] for real-time interaction or edge devices. Mixing multiple variants in the same workflow is not recommended because of different output styles and quality characteristics.
-
FLUX.2 is one of the most important milestones in image generation in 2025-2026. It proves that photorealism and production-level usability can coexist, and it also proves that the open source core model can form a sustainable commercial closed loop in the field of AI imaging. For AI practitioners, designers, and developers, FLUX.2 is a model family that deserves to be taken seriously—whether with its API or running its weights on your own GPU.
User Reviews
-
AnnaOrtiz—FLUX.2 的多参考一致性真的让我震惊，上传 5 张参考图，角色身份保持得特别稳，以前要折腾一晚上的事现在一次生成搞定。 -
Madison_Reyes_Pro_146—用了一段时间 Pro 版本，照片级真实感确实强，人像几乎可以和相机拍出来的混在一起了。不过复杂场景偶尔还会崩，但已经比之前好太多了。 -
Walter.BellIII—价格其实不算贵，$0.03 一张 Pro 图，比 Midjourney 订阅制灵活很多。但如果是大批量生成，成本还是要算一下。 -
Donald.LewisIII—yyds，Klein 版本在 4090 上跑基本秒出图，还开源了 Apache 2.0 协议，商业随便用。BFL 格局是真的大。 -
Bitcoiner680_s—试了下 Flex 版的文本渲染，海报上的中英文小字基本可读，虽然和 Ideogram 3 比起来还差点，但进步已经很大了。 -
Joshua.PerryJr—我想说这个 32K token 的上下文太适合自动化管线了，直接丢 JSON 结构化的提示词进去，它就能理解分层构图。对于批量生成来说是革命性的。 -
LWood062—刚把 ComfyUI 升级了，装上 FLUX.2 的节点，配合 NVIDIA 的 FP8 量化，原来跑不了大模型的 3080 都能跑了，画质还行。 -
ParkerRichter—FLUX.2 的精确颜色控制太香了，直接输 Hex 色号，品牌色零误差。我们电商团队现在批量生成产品图，以前要后期校色，现在一步到位。 -
海角10—Dev 版本的非商业授权有点坑，看标题说「开放权重」就兴冲冲下了，结果商用要 $999/月。团队预算不够，只能是个人项目玩玩。 -
Daniel_Morgan_88—小红书风格图，我用 FLUX.2 Klein 搭配专门的 LoRA，效果比很多在线网站还好。而且本地跑还不用排队，想怎么生成就怎么生成。 -
MrCléoMeunier_pro—和 Midjourney V8 对比了一下，MJ 在绘画风格上依然领先，但论真实感，FLUX.2 确实更强。我的策略是两个都续费，不同场景换着用。 -
angryswan263—看到有人在 HN 上用纯 C 实现了 Klein 的推理引擎，性能虽然还比 Python 栈慢 10 倍，但思路很有意思。开源社区对 FLUX 的热情是真的高。 -
NoahCruz_66433—公司从 Midjourney 切到了 FLUX.2 Pro，主要原因就是 API。MJ 到现在 API 都半残不残的，BFL 的 API 集成起来干净利落。 -
PhilipGutierrezZ0—试了下 4MP 输出，细节确实够丰富，但本地 Dev 版本要 H100 级别才能跑爽，普通 4090 跑量化版还行，全精度就吃力了。 -
greenduck678—说实话在产品摄影这个细分领域，FLUX.2 已经是目前最强的了。我们拍了十几组产品的对比测试，光影和材质表现是最接近真实的。 -
流光518—FLUX.2 的结构化提示词大大减少了试错成本，以前用自然语言调半天，现在直接写 JSON 控制构图，想做一系列风格统一的图直接一套提示词模板搞定。 -
BobbyKing_2022—有没有人发现 Klein 9B 版本在 MacBook Pro M4 Max 上跑也挺流畅的？完全不依赖云端，隐私数据不用担心了。 -
Brenda.Cook_88—64GB VRAM 你认真的吗？跑个图要 60 多 GB 显存，这是要把数据中心搬家里啊。当然量化之后好一些，但门槛还是太高了。 -
Andrew_Chavez_2022—试用下来感觉 PDF/图表生成方面 FLUX.2 是最靠谱的。之前用其他模型生成信息图，文字糊成一团，FLUX.2 的字体清晰很多，基本可以直接用。 -
Lydia_Moore_88—Flex 版一步要 $2.46？太贵了吧，用于高精度场景可以，批量走还是 Pro 合算。 -
Diane_Perez_Pro—BFL 的团队背景确实硬，Stable Diffusion 核心团队出来的，技术路线走得很稳。$430M 融资也不是白拿的，产品迭代速度肉眼可见。 -
Lauren.Stephens_202028—FLUX.2 对国内的适配还不错，中文提示词支持很好，不用硬写英文。而且社区里已经有很多汉化教程了，上手不难。 -
pglnje2sse—感觉 FLUX.2 对光影的理解比上一代强了一个档次，输入「夕阳下的玻璃杯」它真的能算出反射折射效果。以前 AI 画图的「光影扁平」问题基本解决了。 -
ACollins_202289—刚上手时踩了个坑——不知道要登录 Hugging Face 才能下载权重，还以为是网络问题。社区文档确实还不够完善，但跑起来之后效果惊艳。 -
Philip_OrtizJr—Multi-reference 是个好东西，但是超过 8 张图的时候增益就不明显了，而且对 prompt 的自由度有压缩。建议实际使用时控制在 5-6 张。 -
ChloeJackson_eth—FLUX.2 的生态在快速增长，ComfyUI 和 Forge 都已经原生支持了，衍生 LoRA 模型也有大几千个。虽然还不如 SD 生态丰富，但势头很猛。 -
万芳—Adobe 和 Meta 都在用 BFL 的模型，这个信号很明确——企业级图像生成的标准正在被 FLUX 重新定义。 -
魏琪—看了个对比测试，FLUX.2 Pro 和 Nano Banana Pro 各有胜负，但价格只有人家的几分之一。BFL 这个定价策略真的狠。 -
PeterNielsen—作为独立开发者，Klein 的 Apache 2.0 授权对我来说太重要了。不用担心中间商抽成，也不用担心许可证陷阱，直接在项目里集成商用。 -
NAbro—FLUX.2 的 Grounded Generation 功能（web search）有点意思，可以生成实时事件相关的图片。但试了几次有时候搜出来的信息不太相关，还需要打磨。