FLUX 3: Real World Models

Towards Multimodal Flow Models as the Backbone of Visual Intelligence

Date: July 23, 2026 | Read Time: ~5 min | Status: Early Access

---

The Philosophy of World Representation

Black Forest Labs is moving beyond isolated data processing. FLUX 3 is designed to learn from images, audio, and video simultaneously because the goal isn't to master a single medium, but to understand the fundamental nature of reality.

"A model must learn a representation of the world: how objects hold together, how things move, and how events sound."

Each sensory input is essentially a projection of one underlying truth, though each sensor loses specific data:

• Images: Capture spatial relationships and structures at a frozen moment.
• Videos: Reintroduce the temporal dimension, revealing physical laws and dynamics.
• Audio: Uncovers causal links between mechanical actions and acoustics that are invisible to the eye.
• Language: Bridges these perceptions with instructions, abstractions, and intent.

By training on all these simultaneously, the model leverages mutual constraints: modalities are separate $\rightarrow$ modalities are evidence of one reality. For example, the sound must align with the impact, and motion must respect mass.

Technical Foundation: Self-Flow

FLUX 3 is built upon Self-Flow, a proprietary method for aligning multimodal understanding and generation within a single architecture. By scaling compute and data, Black Forest Labs trained the model across video, images, and audio concurrently.

Performance Metrics

The model's efficiency is measured via:

1. Generation Error: Measured by the Fréchet distance, normalized as $FM = 100$ (where $\downarrow$ is better).
2. Manipulation Success: The rate of success in robotic tasks across four distinct groups via finetuning ($\uparrow$ is better).

---

Core Capabilities & Early Evaluations

FLUX 3 can blend modalities, generating joint video and audio from text prompts or using image/video references.

🎥 Video Generation

The model produces diverse clips up to 20 seconds with native audio.

Supported Workflows:

• Text-to-Video: Pure prompt-based generation.
• Image-to-Video: Animation of a starting frame or using images as visual guides.
• Video-to-Video: Transferring elements (like a specific character) into new contexts.
• Audio-Visual Continuation: Extending existing video/audio clips.
• Keyframe-to-Video: Creating controlled transitions between specific moments.
• Agentic Chaining: Linking clips into complex, multi-shot sequences.

Note: The model supports a vast array of aspect ratios and styles, from cinematic to "candid camcorder" footage.

📊 Comparative Benchmarks

In preliminary head-to-head tests (10s, 720p clips with audio), FLUX 3 was preferred over competitors at the following rates:

Strengths: Particularly adept at human facial expressions, multilingual support, and synchronizing sound with physical events.

🖼️ Image Synthesis

FLUX 3 handles image editing and synthesis across various resolutions and styles. Mid-training evaluations show:

• Enhanced Prompt Adherence: Significant leaps in handling complex instructions.
• Typography: High-accuracy text rendering in multiple languages.

---

🤖 Action & Physical AI

The model's understanding of the world extends to action prediction. This is achieved through two primary methods:

1. Native Integration: Scaling Self-Flow to include action prediction directly in the core.
2. Foundation Backbone: Using the video model as a dynamics-aware base for specialized finetuning.

Case Study: FLUX-mimic

In partnership with mimic robotics, the team created FLUX-mimic. This model combines the FLUX 3 backbone with mimic's expertise in dexterous manipulation for production-ready robot learning. This technology is currently being tested on real-world production tasks at Audi.

# Conceptual representation of the Action-Prediction pipeline
def predict_action(visual_state, goal_instruction):
    world_model = FLUX_3_Backbone()
    dynamics = world_model.predict_next_frame(visual_state)
    action = specialized_action_head(dynamics, goal_instruction)
    return action

Launch Roadmap

The rollout will occur over the coming weeks:

• FLUX 3 Video: Currently in Early Access.
• FLUX 3 Image: Early access opening in the following weeks.
• Physical AI Tools: Ongoing integration with partners.