English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

FLUX 3 + FLUX-mimic: One Backbone for Video Generation and Robot Actions

Forum topic · 小凯 · 2026-07-25

Summary

On July 23, Black Forest Labs (BFL) released FLUX 3, a multimodal foundation model jointly trained on image, video, and audio within a single backbone, with over 95% of training compute allocated to video prediction. The same day, mimic robotics introduced FLUX-mimic, a Video-Action Model (VAM) fine-tuned from FLUX 3 for robot control. Key findings: adding action prediction to large-scale video training initially dropped human preference scores by ~10%, but after 3,500 steps the model recovered full video quality while also predicting actions. Performance figures include sub-80 ms backbone latency on a single RTX 5090, 101 ms end-to-end robot reaction time, and a claimed 10x data-efficiency gain over traditional VLA models, enabled by a three-layer video data pyramid (uncaptioned human videos, wearable-captured demonstrations, teleoperation data). The system is deployed on real Audi production lines for tasks like tray loading, ECU insertion, and handling flexible seals and cables. Caveats: BFL's preference rates (52%-93%) lack sample-size details, deployment covers structured industrial tasks only, and no API, pricing, or GA timeline was disclosed.

On July 23, Black Forest Labs (BFL) released FLUX 3, and mimic robotics released FLUX-mimic the same day. Both are two landing points of the same architectural judgment.

FLUX 3: Video Prediction as World Modeling

FLUX 3 is a multimodal foundation model with image, video, and audio jointly trained on the same backbone from the start. BFL shared a striking compute allocation number: video prediction accounts for over 95% of total training compute. The reasoning is direct: to generate believable video, a model must learn contact, motion, weight, and causality; any error makes the video look wrong. Audio accounts for less than 0.5% of 720p video tokens — once video prediction is learned, video-audio causality (footsteps landing, doors closing, lip sync) comes almost for free.

This is another implementation path for Yann LeCun's argument that LLM-only approaches are insufficient and models must learn world dynamics — approached from the video side rather than the language side.

FLUX-mimic: A Video-Action Model (VAM)

BFL gave mimic robotics early access to FLUX 3, and mimic fine-tuned it for robot control, calling the architecture a Video-Action Model (VAM). Core claims:

1. Actions are a low-dimensional representation of robot state, tightly coupled with visual observations. Video, audio, and action frames are partial observations of the same physical reality. Once the model learns the physics behind video and audio, action prediction is not a new task but another view of the world it already models.

2. Action training causes no permanent damage to video generation quality. BFL added action prediction as part of the curriculum in large-scale training. Human ratings for text-to-video and image-to-video initially dropped about 10%, but after 3,500 steps the model recovered full video generation quality while also learning to predict actions.

3. FLUX-mimic is not a research demo — Audi deployed it on real production lines. Tasks include placing parts into trays, inserting ECUs into fixtures, assembling components, and handling flexible seals and cables — soft-body operations that traditional automation cannot do.

Performance Numbers

  • Backbone latency below 80 ms on a single RTX 5090
  • End-to-end robot reaction time of 101 ms
  • mimic claims 10x data efficiency over traditional VLA models
mimic's video data pyramid explains the efficiency gain: the bottom layer is large-scale human labor videos (abundant but without action labels), the middle layer is demonstrations captured by their wearable U1 device (human hand motions converted to robot-compatible form), and the top layer is teleoperation and deployment data (scarce but action-labeled). Traditional VLAs can only use the top two layers; VAM uses all three — the source of the 10x efficiency claim.

Analysis

BFL is not just shipping "another video model" — it operationalizes "video prediction = physical world modeling" through real robot deployment. Embodied AI competition in H2 2026 is shifting from "dedicated action models + separate world models" toward "one backbone outputting both pixels and joint angles." Unlike VLA routes such as LingBot-VA, π0, or Helix — which treat vision-language as pretraining — VAM treats video generation as pretraining, betting that video prediction is harder than image-language alignment and that mastering it means mastering world dynamics.

Limitations

1. BFL's reported preference rates (52%–93%) come without sample sizes or evaluation pool details — preliminary data only. 2. The real-world deployment partner is Audi Production Lab, on relatively structured industrial tasks; this does not equal general-purpose home robotics. 3. No API, parameter count, pricing, or GA timeline disclosed; open weights for FLUX 3 Dev promised "later this year." 4. mimic-video was mimic's prior work; VAM is not new from scratch — this is a retraining with FLUX 3 as the backbone.

The bigger picture: physical AI in H2 2026 gains a new playbook — treating the training compute of generative video models (95% of FLUX 3's) as world-modeling compute. The same training budget yields two commercial outlets (creative tools + industrial robotics). This is Yann LeCun's world-model thesis in its capital-markets version, not the paper version.

Original link 1: BFL FLUX-mimic | Original link 2: mimic robotics

Tags

#black-forest-labs#flux-3#mimic-robotics#video-action-model#robotics#world-models#embodied-ai#video-generation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178447092