Understanding multi-modal generative visual AI

Highlights

• Black Force Labs' Flux model family is positioned as the standard for open image generation, offering Pro, Dev (open source), and Schnell tiers.

• Flux Context unifies text-to-image generation and editing, enabling real-time character consistency, style transfer, and local modifications.

• The core technology, Latent Flow Matching, uses a lower-dimensional latent space and a vector field to efficiently translate from simple to complex image distributions.

• Latent Generative Modeling compresses images to perceptually relevant information, avoiding wasted computation on unperceivable details.

• Flow Matching trains a neural network to map between probability distributions, conditioned on text for generation.

• Flux models achieve orders of magnitude faster generation speeds through Latent Adversarial Diffusion Distillation (LADD), reducing integration steps from 50 to 4.

• The demo showcased Flux Context's ability to perform complex, multi-step image edits—including object manipulation, context changes, and style transfer—while maintaining consistency.

Timeline

1. 00:00 - 03:34

   Introduction to Black Force Labs and Flux Model Family

   Andy, co-founder of Black Force Labs, introduces the company's vision to make visual media the central interface for human communication, providing infrastructure for images and videos. The company, founded in August 2024, has grown to 45 employees with headquarters in Germany and SF. He details the Flux model family, structured into three tiers: Pro (powerful, fast, API-only), Dev (open source, customizable, largest ecosystem on Hugging Face), and Schnell (fully open source, entry-level). Black Force Labs, being the original developers of Stable Diffusion, maintains strong roots in the open-source community, making Flux Dev the standard for open image generation.

2. 03:34 - 08:48

   Flux Context: Unifying Generation and Editing

   The speaker dives into Flux Context, their most recent model, which unifies text-to-image generation and editing, released in June 2025. He emphasizes the critical importance of image editing, which has historically lagged behind generation, for iterating and gaining precise control over existing images. Flux Context combines generation with features like character consistency, style reference, and local editing at near real-time speed. Examples include removing objects, changing context while maintaining character, style transfer, text editing, and solving business problems like generating product shots from in-the-wild images or rendering sketches. This unified approach eliminates the need for time-consuming fine-tuning, significantly speeding up the process of achieving desired results.

3. 08:48 - 18:48

   The Latent Flow Matching Algorithm

   This section explains the core algorithm behind Flux models: Latent Flow Matching, composed of 'Latent' and 'Flow Matching.' The 'Latent' aspect refers to latent generative modeling, which recognizes that much image information is imperceptible to the human eye, making it inefficient to train models in pixel space. Instead, an autoencoder learns a lower-dimensional 'latent space' by compressing images and using a discriminator loss to retain only perceptually relevant details. 'Flow Matching' then operates in this latent space, training a neural network to map a simple standard normal distribution to the complex data distribution of natural images, conditioned on text input. This process involves integrating along trajectories represented by the neural network to generate images from text prompts.

4. 18:48 - 26:32

   Flux Context Architecture and Speed Optimization

   The speaker details the Flux Context architecture, which is a transformer model designed for both text-to-image generation and editing. Its input includes text tokens and visual tokens, with separate sets of visual tokens for generation and context images during editing. The model uses special 'double stream blocks' for handling visual and text tokens separately, except during joint attention operations. A key innovation for speed is Latent Adversarial Diffusion Distillation (LADD), which drastically reduces the numerical integration steps from around 50 to just 4, resulting in a 12.5x speedup. LADD simplifies the original Adversarial Diffusion Distillation by moving all operations to the latent space and using the teacher model as a discriminator, significantly improving memory efficiency and training effort.

5. 26:32 - 31:22

   Live Demo and Conclusion

   The presentation concludes with a live demo showcasing Flux Context's capabilities in image editing. Starting with an SC Freiburg logo, the speaker demonstrates putting the logo onto a t-shirt, then iteratively making the logo smaller and changing the t-shirt color to red, all while maintaining object consistency. He then performs a more complex transformation by placing the t-shirt onto a man walking in a park, highlighting the model's ability to handle significant context changes. Finally, a style transfer is applied, turning the image into a watercolor painting, illustrating the model's versatility in combining multiple tasks. The demo underscores the power and generality of Flux Context, which can perform tasks that previously required extensive fine-tuning or were impossible with traditional tools.