Black Forest Labs has released weights for a model that generates images from text and edits reference photos. Developers building image tools on their own hardware gain an option, but must consider the prompt expansion model and usage terms alongside the smaller generator.

1. Combining generation and editing

Released on October 15, 2026, the model combines up to 10 reference images and supports resolutions up to 4K.

The 5.6 billion parameter count covers only the image generator, excluding the text encoder that converts text into model-readable representations.

There is a base model for further training and a distilled model for generation and editing. Distillation reduces the computational steps needed for generation; the base model uses 50 steps and the distilled model uses 16.

2. Quality and speed depend on measurement scope

The developer reported first place in internal evaluations of text-to-image generation, multi-reference generation, and editing. Comparators Qwen-Image 2.1 and Krea 2 have generators with 7 billion and 12.9 billion parameters, respectively. Krea 2 lacks editing support and was excluded from that comparison. No independent evaluation results were provided.

NVIDIA measured GPU times using the fastest build for each device, while consumer card results came from ComfyUI. The results therefore do not compare identical software configurations.

For one 1024×1024 image, the distilled model's 16-step generation takes 0.55 seconds on GB300 and 8.9 seconds on RTX 4070.

In 50-step measurements of the base model at the same resolution, GB300 took 1.7 seconds and DGX Spark took 43.8 seconds.

Median API latency at 1 megapixel is 4.54 seconds for generation and 7.18 seconds for editing. It includes queuing, networking, and prompt expansion, so it should not be compared with GPU time. Measurement boundaries also differ across products: its own models were measured server-side, while competing models were measured client-side.

3. A smaller generator does not eliminate total memory costs

Self-hosting and further training are easier to consider, but some costs shift to input processing. By default, the reference code runs a separate 9 billion-parameter model that expands short prompts into scene descriptions covering lighting and composition. The model card recommends this expansion step, saying that short prompts sent directly produce generic results.

In the reference code, peak GPU memory for the distilled model on B200 was 31GB in the default 16-bit format at 1024×1024 with prompt expansion disabled. Quantized formats, which store values at lower numerical precision, used 26GB for FP8 and 25GB for NVFP4. The expansion model uses about 28GB more on an 80GB GPU.

Downloading the fast model and prompt expander requires 52GB of storage; downloading both models and all related files requires 107GB. Download size and runtime memory are separate, so generator parameter count alone is not enough to choose hardware.

4. Choose the execution path first

After installing the reference repository and downloading the weights, generate an image with the following command. This is the original example, with prompt expansion enabled by default.

python scripts/cli.py "A man walking with his dog, afternoon sun" --model distilled --out out.png --seed 7

For 8GB GPUs, guidance points to INT8 files for ComfyUI. The reference code has no execution path for this precision, so support for smaller hardware should not be confused with the reference code's memory figures.

5. Check editing boundaries and licensing separately

Bounding boxes specifying element positions and sizes are not masks that limit changes to a cropped region. Elements may extend beyond boxes, small elements may be omitted, and areas outside the edit region may change. The API's exact preservation of unedited areas results from a separate restoration feature.

The model card states that text in images may be inaccurate or distorted and that training data biases may appear. Outputs are not designed to provide factual information.

The weights' FLUX Kommunity License v.1.0 conditionally permits commercial output generation for eligible users with annualized gross revenue, including affiliates, below $5 million. It does not broadly permit commercial further training or distribution. Restrictions also cover military use, surveillance, biometric processing, and improving other models with similar functionality.

Adoption should depend less on the generator's size than on whether runtime costs, including prompt expansion, and licensing fit the task.