Black Forest Labs releases FLUX.2 [klein]a compact picture mannequin household focusing on interactive visible intelligence on client {hardware}. Flux.2 [klein] It extends the FLUX.2 line with sub-second era and modifying, built-in text-to-image and image-to-image architectures, and deployment choices starting from native GPU to cloud APIs, whereas sustaining state-of-the-art picture high quality.
From FLUX.2 [dev] In the direction of interactive visible intelligence
Flux.2 [dev] is a 32 billion parameter rectified circulation transformer for textual content conditional picture era and modifying, together with composition with a number of reference photographs, and runs totally on datacenter-class accelerators. Tuned for optimum high quality and adaptability, with lengthy sampling schedules and excessive VRAM necessities.
Flux.2 [klein] takes the identical design course and compresses it into small rectifier transformers with 4 billion and 9 billion parameters. These fashions are tuned to very quick sampling schedules, assist the identical text-to-image modifying duties, multi-reference modifying duties, and are optimized for sub-second response occasions on trendy GPUs.
Mannequin household and options
Flux.2 [klein] The household consists of 4 primary open weight variants by way of a single structure.
- Flux.2 [klein] 4B
- Flux.2 [klein] 9B
- Flux.2 [klein] 4B base
- Flux.2 [klein] 9B base
Flux.2 [klein] 4B and 9B are the step distillation mannequin and the steerage distillation mannequin. They use 4 inference steps and are positioned because the quickest possibility for manufacturing and interactive workloads. Flux.2 [klein] 9B combines the 9B circulation mannequin and the 8B Qwen3 textual content embedder and is alleged to be the flagship compact mannequin on the Pareto frontier of text-to-image high quality and latency, single-reference modifying, and multiple-reference era.
The Base variant is an undistilled model with an extended sampling schedule. The documentation lists these as base fashions that protect the whole coaching sign and supply increased output variety. These are meant for fine-tuning, LoRA coaching, analysis pipelines, and customized post-training workflows the place management is extra vital than minimal delay.
All FLUX.2 [klein] The mannequin helps three core duties with the identical structure. You may generate photographs from textual content, edit a single enter picture, or carry out multi-reference era and modifying the place a number of enter photographs and prompts work collectively to outline a goal output.
Latency, VRAM and quantized variants
Flux.2 [klein] The mannequin web page supplies approximate end-to-end inference occasions on GB200 and RTX 5090. FLUX.2 [klein] 4B is the quickest variant, listed at roughly 0.3 to 1.2 seconds per picture, relying in your {hardware}. Flux.2 [klein] 9B is top of the range and targets about 0.5 to 2 seconds. The bottom mannequin runs on a 50-step sampling schedule, which takes just a few seconds, however permits extra flexibility for customized pipelines.
Flux.2 [klein] The 4B mannequin card states that 4B matches round 13 GB of VRAM and is appropriate for GPUs such because the RTX 3090 and RTX 4070. [klein] The 9B card stories a VRAM requirement of roughly 29 GB and targets {hardware} such because the RTX 4090. Which means a single high-end client card can host distilled variants with full decision sampling.
To develop software to extra units, Black Forest Labs may even launch FP8 and NVFP4 variations for all FLUX.2. [klein] Variant co-developed with NVIDIA. FP8 quantization is alleged to scale back VRAM utilization by as much as 40 p.c and be as much as 1.6x quicker on RTX GPUs whereas sustaining the identical core performance, and NVFP4 is alleged to scale back VRAM utilization by as much as 55 p.c and be as much as 2.7x quicker on RTX GPUs.
Benchmarking in opposition to different picture fashions
Black Forest Labs evaluates FLUX.2 [klein] By way of Elo type comparability of textual content and pictures, single reference modifying, and a number of reference duties. Efficiency chart exhibits FLUX.2 [klein] Concerning the Pareto frontier of Elo rating vs. latency and Elo rating vs. VRAM. The outline says FLUX.2. [klein] It matches or exceeds the standard of Qwen-based picture fashions at a fraction of the latency and VRAM, and supplies higher efficiency than Z Picture whereas supporting unified text-to-image conversion and a number of reference modifying in a single structure.
The fundamental variant trades some velocity for full customizability and fine-tuning. That is in keeping with its function as a foundational checkpoint for brand spanking new analysis and domain-specific pipelines.
Vital factors
- Flux.2 [klein] is a compact rectified circulation transformer household with 4B and 9B variants that helps text-to-image conversion, single picture modifying, and a number of reference era in a single built-in structure.
- Distilled FLUX.2 [klein] The 4B and 9B fashions use 4 sampling steps and are optimized for subsecond inference on a single trendy GPU. However, the undistilled Base mannequin makes use of an extended schedule and is meant for fine-tuning and analysis.
- Quantized FP8 and NVFP4 variants constructed with NVIDIA ship as much as 1.6x speedups with roughly 40 p.c much less VRAM for FP8 on RTX GPUs, and as much as 2.7x speedups with roughly 55 p.c much less VRAM for NVFP4.
Please test technical details, lipo and model weights. Please be at liberty to comply with us too Twitter Do not forget to hitch us 100,000+ ML subreddits and subscribe our newsletter. hold on! Are you on telegram? You can now also participate by telegram.
Michal Sutter is a knowledge science skilled with a grasp’s diploma in knowledge science from the College of Padova. With a powerful basis in statistical evaluation, machine studying, and knowledge engineering, Michal excels at remodeling advanced datasets into actionable insights.
