A new era of image editing has arrived with the introduction of the FLUX.1 Kontext [dev] model by Black Forest Labs. Building upon thier Flow.1 context family of image models,this latest innovation empowers users to manipulate images with unparalleled precision and ease,utilizing both text and image prompts.

Unlike traditional methods that require intricate instructions and complex masks,the FLUX.1 Kontext [dev] model offers a guided, step-by-step generation process, granting users granular control over every aspect of image evolution. Whether refining minute details or fully transforming a scene, this open-weight generative model ensures coherent, high-quality edits that remain faithful to the original concept.

Key capabilities of FLUX.1 Kontext include:

• Character Consistency: Maintain unique character traits across diverse scenes and angles.
• Localized Editing: Precisely modify specific elements without affecting the surrounding areas.
• Style Transfer: Seamlessly apply the aesthetic of a reference image to new creations.
• Real-Time Performance: Experience low-latency generation for rapid iteration and immediate feedback.

The weights for FLUX.1 Kontext are available for download on hugging Face, along with TensorRT-accelerated variants.

The [dev] model prioritizes adaptability and control, offering features like character consistency, style preservation, and localized image adjustments, enhanced by integrated ControlNet functionality for structured visual prompting.

The Flow.1 context [dev] is currently accessible in ComfyUI and the Black forest Labs Playground, with an NVIDIA NIM microservice version anticipated in August.

NVIDIA RTX Optimization Through TensorRT Acceleration

The collaborative effort between NVIDIA and Black Forest Labs has resulted in a model that not only simplifies complex workflows but also broadens accessibility. By quantizing the model and optimizing it with TensorRT, they have achieved a meaningful reduction in VRAM requirements and a doubling of performance.

Quantization reduces the model size from 24GB to 12GB for FP8 (Ada) and 7GB for FP4 (Blackwell). The FP8 checkpoint is optimized for GeForce RTX 40 Series gpus,leveraging their FP8 accelerators. The FP4 checkpoint is tailored for GeForce RTX 50 Series GPUs, utilizing a novel method called SVDQuant to maintain image quality while minimizing model size.

TensorRT, a framework designed to harness the power of Tensor Cores in NVIDIA RTX GPUs, delivers over 2x acceleration compared to running the original BF16 model with PyTorch.

“This enables coherent, high-quality image edits that stay true to the original concept.”

For detailed information on NVIDIA optimizations and guidance on using FLUX.1 Kontext [dev], refer to the NVIDIA Technical blog.

Getting Started with FLUX.1 Kontext

The Flow.1 context [dev] model is readily available on Hugging Face in both Torch and TensorRT formats.

AI enthusiasts can experiment with the Torch variants in ComfyUI. Additionally, Black Forest Labs offers an online playground for convenient model testing.

For advanced users and developers, NVIDIA is developing sample code to facilitate seamless integration of TensorRT pipelines into existing workflows. The DemoDiffusion repository will be available later this month.

additional AI Innovations

google recently unveiled GEMMA 3N,a new multimodal small language model optimized for NVIDIA GeForce RTX GPUs and the Nvidia Jetson platform.

Users can leverage Gemma 3n models with RTX accelerations in Ollama and Llama.cpp with applications like AnythingLLM and LM Studio.

Developers can easily deploy Gemma 3n models using Ollama and benefit from RTX accelerations.Instructions for running Gemma 3n on Jetson and RTX are available.

Furthermore, NVIDIA’s Plug and Play: Project G-Assist Plug-In Hackathon, a virtual event concluding on Wednesday, July 16, invites developers to create custom G-Assist plug-ins for a chance to win prizes. A G-Assist Plug-In webinar is scheduled for Wednesday, July 9, from 10-11 a.m. PT.

Join NVIDIA’s Discord server to connect with community developers and AI enthusiasts.