Black Forest Labs Releases FLUX.3 Image Klein

Black Forest Labs Releases FLUX.3 Image Klein

At a Glance

Item Value
Repository black-forest-labs/flux-3-image-klein-pup
Family guide flux-3-image-klein guide (1 articles)
Publisher guide Black Forest Labs: models and licenses
Published 2026-10-13
License flux-kommunity-license
Formats safetensors
Access Gated on Hugging Face (license agreement required)
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Black Forest Labs has released a new open image generation model, “FLUX.3 Image Klein". It is supported from day one in Hugging Face’s Diffusers library.

A major feature of this model is its ability to handle not only text-to-image generation, but also image-to-image generation and editing using multiple reference images, as well as layout control using bounding boxes, all within a single model. By including bounding boxes specifying regions within the prompt, users can finely specify where subjects are placed in the image, or where specific parts of reference images should be moved.

Two types of checkpoints have been released: “flux-3-image-klein", which is distilled for fast inference, and “flux-3-image-klein-base", a base model intended for uses such as fine-tuning.

Specifications

The main specifications and recommended settings described in the published materials are as follows:

  • Architecture: Employs a latent-space flow matching model. The transformer consists of two stages. In Stage 1, the text, noisy image being generated, and reference image streams pass through five per-mode blocks with independent weights. In Stage 2, the concatenated sequence passes through 28 joint blocks with shared weights. Within each block, attention and a SwiGLU MLP operate in parallel.
  • Text Encoder: Employs the vision-language model “Qwen3-VL". Rather than using only the final layer, hidden states from 8 out of all 36 layers (layers 4, 8,…, 32) are stacked to generate a 20,480-dimensional vector, which is projected via a linear layer to the transformer width of 3,072 dimensions.
  • VAE: Employs the same VAE architecture as FLUX.2. Images are compressed to 1/8th of their height and width into 32 channels, and further grouped into 2×2 patches into a 128-channel single token (equivalent to 16×16 pixels per token).
  • Positional Encoding: Employs 4D Rotary Position Embedding (RoPE) with four axes: timestep (t), height (h), width (w), and sequence length (l).
  • Output Specifications: Default resolution is 1024×1024 pixels. Height and width can be changed when running scripts.
  • Recommended Inference Steps, Sampler, and CFG:
  • Distilled version (flux-3-image-klein): Recommended 16 steps, Classifier-Free Guidance (CFG) is not used. – Base version (flux-3-image-klein-base): Recommended 50 steps, supports CFG. – A large time shift (6.93) is set for the scheduler, and 44 out of 50 steps in the base version run in the high-noise region (σ > 0.5).
  • Distribution Formats and Variants: In addition to bfloat16, single-file quantization model formats including row-wise FP8 (fp8r) and NVFP4 (nvfp4) are available.
  • Licensing Conditions: Both models are subject to the “flux-kommunity-license". The Hugging Face repository is a gated repository that requires agreement to the terms of use.

Performance and Quality

Official blog posts from the publisher report measured generation speeds for the distilled model (16 steps) using an NVIDIA H100 80GB (without offloading, default attention backend). The averages of 3 runs after warm-up are as follows:

  • Text-to-image (1024×1024 generation): 1.93 seconds
  • Image-to-image (one 1024×1024 reference image, 1024×1024 generation): 3.87 seconds
  • Layout editing (one 1360×768 reference image, 1360×768 generation): 4.10 seconds

In workflows using reference images, tokens from the reference images are integrated into the same sequence as the text and generated image, adding 4,096 tokens per reference image of 1024×1024 per transformer call. As a result, processing time is reported to be approximately double compared to text-to-image. The number of tokens per transformer call is 4,256 tokens for text-to-image, 8,272 tokens for image-to-image, 4,336 tokens for layout control text-to-image, and 8,800 tokens for layout editing.

Note that there are no quantitative comparison benchmarks (such as VBench score tables) against existing models included in the published materials.

Strengths and Use Cases

FLUX.3 Image Klein can handle three roles in a single model: text-to-image generation, image editing using multiple reference images, and precise layout control using bounding boxes. In particular, the ability to control elements within an image using pixel-level coordinate specification is a major strength not found in conventional models.

In layout control, a list of regions in JSON format is appended after the prompt. Each region describes an ID, bounding box ([y_min, x_min, y_max, x_max] format, on a scale of 0 to 1000), and a description of that location. This enables advanced editing, such as specifying subject placement in a new image to the millimeter or moving specific elements within a reference image (e.g., a knitted knight) to another position while maintaining the background.

To maximize the performance of this model, detailed scene descriptions covering subjects, settings, lighting, composition, and style are recommended. To assist with prompt creation, a dedicated model called “FLUX 3 Image Klein Prompt Upsampler (PUP)" is provided. This is a 9B parameter vision-language model based on Qwen3.5-9B, fine-tuned to expand short requests into detailed prompts. Since inputting short prompts directly tends to yield generic outputs, a workflow placing this PUP in the pre-stage is recommended.

Expected use cases include generating photorealistic photographs, graphic design requiring complex compositional specifications, and advanced image editing to accurately move or alter parts of existing images. The base model also supports LoRA fine-tuning, making it suitable for training adapters specialized in specific styles (such as tarot card aesthetics) or instruction-based image editing.

How It Differs from Similar Models

Models based on the same Qwen3.5-9B foundation include “clef-flash" structured output-focused model: example answers on this site, required VRAM 4GB+, GGUF version available and “LensVLM-9B" long-text reading vision-language model that compresses images for processing: required VRAM 4GB+, GGUF version available. While these are primarily aimed at “understanding" or “reasoning" about visual information, the decisive difference is that FLUX.3 Image Klein applies those insights to image “generation" and “editing", specializing in accurately outputting images according to instructions.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
8GB (RTX 4060 / 3060 Ti, etc.) Q4_K_XL 5.6GB 6.7GB
12GB (RTX 4070 / 3060 12GB, etc.) Q8_0 8.9GB 10.6GB
24GB (RTX 4090 / 3090, etc.) BF16 16.7GB 20.0GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build unsloth/flux-3-image-klein-pup-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — flux-kommunity-license: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Access: You must accept the license on Hugging Face before downloading.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (flux-kommunity-license). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Distributed Files

Weight files published in black-forest-labs/flux-3-image-klein, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
ae.safetensors 336MB
model.safetensors 11.10GB
text_encoder/model-*-of-00002.safetensors 8.88GB (2 split files)
transformer/diffusion_pytorch_model-*-of-00002.safetensors 11.10GB (2 split files)
vae/diffusion_pytorch_model.safetensors 336MB
variants/fp8r/model.safetensors 6.27GB
variants/nvfp4/model.safetensors 5.28GB

How to Get It

Models in the FLUX.3 Image Klein series are available from Hugging Face repositories. However, because the image generation model itself is a gated model requiring license agreement for access, reviewing and consenting to the terms on Hugging Face is required before downloading.

The distribution format is safetensors, and the following repositories have been published:

  • Image Generation and Editing Models: black-forest-labs/flux-3-image-klein (distilled version), black-forest-labs/flux-3-image-klein-base (base version)
  • Prompt Upsampling Model: black-forest-labs/flux-3-image-klein-pup (this is under the Apache 2.0 license)
  • Single File for ComfyUI: black-forest-labs/flux-3-image-klein-comfy (ComfyUI templates are also provided)

When using the Diffusers library, after installing the latest diffusers and transformers, load using ModularPipeline. For the base model, quantization variants such as row-wise FP8 (fp8r) and NVFP4 (nvfp4) are also available and can be loaded via the from_single_file method. To save memory, methods such as executing the text encoder separately to free up memory or using enable_auto_cpu_offload to move components to the GPU component by component are recommended.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-17 unsloth GGUF unsloth/flux-3-image-klein-pup-GGUF Q4_K_XL 6.7GB (fits in 8GB VRAM)
2026-11-17 mlx-community MLX mlx-community/flux-3-image-klein-pup-4bit MLX 4bit 6.6GB (fits in 8GB VRAM)
2026-11-17 mlx-community MLX mlx-community/flux-3-image-klein-pup-8bit MLX 8bit 11.7GB (fits in 12GB VRAM)

File sizes of each build:

  • Available builds in unsloth/flux-3-image-klein-pup-GGUF: Q4_K_M 5.2GB / Q4_K_XL 5.6GB / Q6_K 6.9GB / Q8_0 8.9GB / BF16 16.7GB
  • Available builds in mlx-community/flux-3-image-klein-pup-4bit: MLX 4bit 5.5GB
  • Available builds in mlx-community/flux-3-image-klein-pup-8bit: MLX 8bit 9.7GB

In addition, 15 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-17: Added converted builds to “Quantized and Converted Variants”: unsloth/flux-3-image-klein-pup-GGUF, mlx-community/flux-3-image-klein-pup-4bit, mlx-community/flux-3-image-klein-pup-8bit
  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of unsloth/flux-3-image-klein-pup-GGUF.