FLUX 3 — Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence

TL;DR

FLUX 3 is Black Forest Labs’ new multimodal foundation model, a single flow-matching architecture trained jointly on images, video, and audio and extended to action prediction. It scales up BFL’s earlier Self-Flow recipe (Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis) so the same backbone drives content-creation (image / video+audio generation and editing) and physical AI (video-action models such as FLUX-mimic, filed separately as FLUX 3 x mimic: The Next Generation of Video-Action Models). FLUX 3 Video generates up to 20 s of 720p video with native audio in one shot, reports pairwise-preference wins over Grok Imagine Video (69%), Kling v3 Pro (60%), Happy Horse 1.1 (57%), Seedance 2.0 and Gemini Omni Flash (~52%), Runway Gen-4.5 (77%), and Luma Ray 3.2 (93%). BFL frames the model as the point at which “visual intelligence” and physical-AI stacks stop needing separate foundation models.

Key claims

FLUX 3 is one flow-matching model jointly trained across image, video, and audio, built on Self-Flow, and extended natively to action prediction; content creation and action prediction are treated as two applications of the same backbone [§Capabilities & Early Evaluations].
Video generation covers text-to-video, image-to-video (animation and reference), video-to-video, generative video-audio continuation, keyframe-to-video, multilingual dialogue, high style diversity, strong typography, and agentic chaining of multi-shot sequences, all with native audio [§Video].
Preliminary pairwise-preference numbers on 10-second 720p T2V-with-audio clips: FLUX 3 preferred over Grok Imagine Video in up to 69% of comparisons, Kling v3 Pro 60%, Happy Horse v1 59%, Happy Horse 1.1 57%, Seedance 2.0 and Gemini Omni Flash 52%, Runway Gen-4.5 77%, Luma Ray 3.2 93% [§Video].
FLUX 3 supports single-generation clips of up to 20 seconds with audio and can be chained into multi-minute sequences with visual references maintaining character consistency across scenes [§Video].
Image capabilities include improved complex-prompt handling and multilingual high-accuracy text rendering; FLUX 3 Image will follow FLUX 3 Video in early access [§Image].
Two paths to action prediction: (i) native action prediction integrated into FLUX 3 directly, scaling up the Self-Flow initial work, and (ii) using the pretrained video backbone as a dynamics-aware foundation that specialized action models are fine-tuned from (the FLUX-mimic route with mimic robotics) [§Action].
Launch plan: video+audio generation and editing via API + private weight access (FLUX 3 Video); action prediction via research/commercial partners starting with mimic robotics (FLUX-mimic, FLUX 3 Action); image synthesis + editing via API + private weight access (FLUX 3 Image); and an open-weights multimodal backbone for content creation and action prediction (FLUX 3 Dev) [§Launch Plan].

Method

FLUX 3 is presented as the scaled-up incarnation of Self-Flow (Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis), BFL’s earlier proposal to unify multimodal generation and representation learning inside a single flow-matching framework. The stated design principle is that no single modality gives a complete description of the world — images capture spatial structure, video adds time and physical dynamics, and audio captures causal relationships between mechanical events and their acoustic consequences — so joint training under mutual constraints (“the sound has to match the impact, the motion has to obey the mass, the future has to follow from the past”) is what turns modality streams into shared evidence about one physical reality [§FLUX 3: One model, multiple capabilities.].

Architecturally the post keeps details private: a single unified architecture, flow matching, trained jointly on image + video + audio at significantly scaled compute and data compared to Self-Flow, with action prediction integrated as an additional modality of the same backbone. The companion post (FLUX 3 x mimic: The Next Generation of Video-Action Models) supplies the missing quantitative detail — that video prediction dominates the training bill at over 95% of total compute, that audio is <0.5% of tokens in a 720p video with audio, and that adding action prediction to the curriculum causes a temporary ~10% drop in human-rated video quality that recovers in ~3500 steps.

Results

Reported evaluations are explicitly preliminary and are gathered on 10-second 720p T2V-with-audio clips during the current early-access phase [§Video]. Headline pairwise-preference numbers vs contemporary systems: 69% over Grok Imagine Video, 60% over Kling v3 Pro, 59% over Happy Horse v1, 57% over Happy Horse 1.1, 52% over Seedance 2.0, 52% over Gemini Omni Flash, 77% over Runway Gen-4.5, and 93% over Luma Ray 3.2 [§Video]. BFL flags FLUX 3 Video as particularly strong on human facial expression, sound-event association, and multilingual dialogue, with visual-reference conditioning used to maintain character consistency across multi-scene sequences longer than the 20 s single-generation limit [§Video]. No image-side numbers, no action-side numbers on standard robotics benchmarks, and no direct comparison against Veo 3 are included in this post; the FLUX-mimic sibling post supplies the action-side evaluation.

Why it’s interesting

The framing collapses several axes the wiki has been tracking as separate races: it is simultaneously (i) an entry in the Joint audio-video generation cluster (jointly-trained T2AV up to 20 s, with pairwise wins over Kling v3 Pro / Seedance 2.0 that are the current open+closed frontier of that concept), (ii) an entry in the World Foundation Models cluster asserting the “content creation and physical AI share one backbone” thesis via the video-95%-of-compute argument, and (iii) an entry in the Unified Multimodal Models cluster with a fourth modality (action) integrated natively rather than bolted on. It is also BFL’s public follow-through on the roadmap tease from Black Forest Labs — 'visual intelligence and the open infrastructure that powers it' as the next step (“visual intelligence and the open infrastructure that powers it”) and the direct scale-up of their own Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis research — a rare case where a lab’s own research paper and its next flagship model are the same recipe at different budgets. Companion post FLUX 3 x mimic: The Next Generation of Video-Action Models carries the physical-AI half of the story.