Flux 3
TL;DR Highlight
FLUX 3 learns images, video and audio in one model; video enters early access
Who Should Read
Backend or product developers comparing video and image generation APIs for their service
Core Mechanics
- FLUX 3 trains on three modalities together
- Images, video and audio are learned in one unified architecture.
- Each modality is framed as a different sensor view of the same reality.
- Video runs up to 20 seconds with audio in one generation
- It supports text-to-video, image-to-video, video-to-video and keyframe-to-video.
- All outputs come with native audio.
- Early preference tests favor FLUX 3 over most rivals
- Based on 10-second 720p text-to-video clips with audio.
- Preferred over Runway Gen-4.5 in 77%, Luma Ray 3.2 in 93%, Kling v3 Pro in 60%, and Seedance 2.0 and Gemini Omni Flash in 52%.
- Clips can be chained into multi-minute sequences
- An agentic flow chains individual clips into multi-shot sequences.
- Visual references keep characters consistent across scenes.
- Robot action prediction uses the same backbone
Community Reactions
- Open weights are the key concern
- Commenters say Flux 2 Dev Klein is the best option on common hardware, so a follow-up matters.
- Others hope VRAM requirements drop versus Flux 2.
- Demo skepticism
- Commenters note almost no examples of people and call the 'World Model' label frivolous.
- They also say the 20-second claim is shown only with jump cuts.
- Early users are impressed
- One commenter saw early-access clips and reflections and expects it to be the new SOTA for home use.
- Another thought the clips were real footage.
- Image quality still questioned
- One commenter says Nano Banana and GPT-2 Image remain best in class and Flux doesn't cross that bar.
How to Apply
- Compare against your current video model with identical prompts
- Request early access and prepare 5 representative prompts.
- Put results side by side and judge audio sync and facial expressions.
- Tabulate the rollout stages when shortlisting models
- List API and open-weight availability for Video, Image, Action and Dev.
- Mark requirements needing open weights as on hold until Dev ships.
- Design a reference-based character consistency pipeline
- Fix a character reference image and feed it into video-to-video calls.
- Split clip stitching into a separate workflow step.
Try It
# Paste into Claude or ChatGPT (framework for comparing video generation models)
I want to compare text-to-video models {{ModelA}} and {{ModelB}}. My use case is {{use case}}.
Create 5 test prompts (10 seconds, 720p) and a scoring sheet using these criteria:
- Facial expressions
- Sound matching physical events
- Multilingual dialogue
- Text (typography) rendering
- Character consistency
Next step: after I give you both outputs, mark which one is preferred per criterion.Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.