Black Forest Labs/FLUX.3 Image Klein 5.6B

• Original model: black-forest-labs/flux-3-image-klein
• License: FLUX Kommunity License. Review its non-commercial terms and qualifying-user commercial-output exception before use.
• Quantization: Abiray/flux-3-image-klein-sd-GGUF, distilled Q4_K_M checkpoint.

FLUX.3 Image Klein combines text-to-image generation and reference-image editing. The 5.6B parameter count describes the diffusion transformer, not the separate Qwen3-VL-4B text encoder. This entry is the sixteen-step distilled model, not Klein Base or the optional prompt upsampler.

Usage

Build the updated image engine with just engine image cpu (or the matching CUDA variant), then serve flux3-klein:5.6b with self serve flux3-klein:5.6b. Existing installed sd-server binaries must be rebuilt: the old master-929-3f8527a engine does not implement FLUX.3.

Generation and editing use the existing OpenAI-compatible image endpoints. Defaults are 16 Euler steps, CFG 1.0, flash attention and CPU offloading. Upstream supplies the model-specific time shift. Although the model supports up to ten reference images, this server caps edits at eight.

Runtime and Validation

Experimental integration using upstream stable-diffusion.cpp PR #2143, pinned to 6d6ff751882e80e8f3af0942c9d4eb4e07375b19. The PR is not yet merged. It implements FLUX.3 architecture detection, Qwen3-VL layer conditioning and reference-image editing; FLUX.2 runtime support alone is insufficient.

The text encoder is used through --llm; this pipeline does not require the Qwen vision projector used by Qwen Image 2.1. The VAE is shared with FLUX.2. All files have immutable revision, byte-size and SHA-256 pins from Hugging Face LFS metadata. The diffusion GGUF is a community conversion, not an official BFL artifact. Original BFL repositories require license acceptance.

Actual weight loading, visual quality, CPU/CUDA inference and reference editing have not been verified through self. Do not treat unit tests or an engine build as an inference smoke test. No performance or quality benchmark is claimed.

Artifacts

Role	File	Bytes	SHA-256
Diffusion	flux3_image_klein_sd_Q4_K_M.gguf	3269858080	c0dc60488f1d73517e60252d0d1ce24a8a9a19275ea669c0b96408a10fc871ac
VAE	split_files/vae/flux2-vae.safetensors	336211292	868fe7b343cc8f3a19dbcfcafbc3d5f888802be3f89bd81b65b3621a066ce8f3
Text encoder	Qwen3VL-4B-Instruct-Q4_K_M.gguf	2497281664	66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a

Registry details

• Registry id: flux3-klein:5.6b
• Type: image
• Default quantization: q4
• Capabilities: text, image, multi-image -> image, image-edit

Benchmarks

No submitted benchmark reports yet.

Available quants

Quant	Adapter	Repository	File	Size	SHA-256
q4 (default)	sd-server	Abiray/flux-3-image-klein-sd-GGUF	flux3_image_klein_sd_Q4_K_M.gguf	3.0 GiB	c0dc60488f1d73517e60252d0d1ce24a8a9a19275ea669c0b96408a10fc871ac
			split_files/vae/flux2-vae.safetensors	321 MiB	868fe7b343cc8f3a19dbcfcafbc3d5f888802be3f89bd81b65b3621a066ce8f3
			Qwen3VL-4B-Instruct-Q4_K_M.gguf	2.3 GiB	66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a