XNews

Black Forest Labs Releases FLUX.2 and Model Testing

Black Forest Labs Releases FLUX.2 and Model Testing

- Black Forest Labs has released the open-source image-to-image and image editing model FLUX.2. - It supports referencing up to 10 images simultaneously. - Offers richer details, clearer textures, and stable lighting. - Enhances capabilities for complex layouts, infographics, and text rendering. - Supports image editing features up to 4MP. - Four model versions have been released: pro, flex, dev, klein (coming soon) along with a new VAE. - The pro version claims to be on par with the best closed models while being faster and more cost-effective. - The flex version provides controllable parameters for steps and guidance strength. - The dev version is a 32B open-weight base model for community development. - The klein version is a distilled open-source model of the same size (Apache 2.0). - User testing: The official 2K version performs well in consistency, but there are differences in multimodal knowledge and consistency compared to Banana. - There were cases of generated missing objects during testing, and compliance with Chinese prompts still has issues.

Key insights

Official Release and Version Composition

Black Forest Labs has released the open-source model FLUX.2, which includes four versions and one VAE: FLUX.2 [pro], [flex], [dev], the upcoming [klein], and the new FLUX.2 - VAE. The official description includes open-source weights and different positioning: pro is positioned to compete with the best closed models, flex offers controllable parameters, dev is a 32B open-weight base model for community use, klein is a distilled open-source model of the same size, and VAE is used for latent representation and quality/compression balance.

Performance, Speed, and Cost Claims

The official claims that FLUX.2 [pro] matches the best closed models in image quality, prompt compliance, and visual realism, while generating faster and at a lower cost. FLUX.2 [flex] allows developers to control steps and guidance strength to adjust between quality, prompt compliance, and speed. FLUX.2 [dev] provides 32B open weights for community use and secondary development.

User Testing: Consistency and Detail Performance

Multiple user tests show significant improvements in detail, texture, and lighting stability with FLUX.2. The official 2K version performs better in consistency than some competitors. Comparative tests show that FLUX.2 performs well in color and material control, text rendering, and complex layouts, but there are differences in multi-image consistency compared to models like Banana (with some cases generating missing objects), and comparisons with Midjourney show varying strengths in different scenarios.

Designer-Friendly and Editing Capabilities

FLUX.2 supports referencing up to 10 images simultaneously, providing more precise color value control and image editing capabilities up to 4MP, making it suitable for designers for multi-input editing and consistency control. It has also improved in following complex, structured instructions and text rendering.

Limitations and Future Potential

Tests and discussions indicate that FLUX.2's LLM/multimodal reasoning is not as strong as some alternative models in real-world knowledge and multimodal reasoning. Compliance with Chinese prompts still has issues in certain scenarios. Authors and testers suggest switching VLMs or improving the prompt process to enhance stability with Chinese and complex prompts.

Source posts