FLUX 3 Image Klein (Step-Distilled) - GGUF

GGUF quantizations of Black Forest Labs' FLUX 3 Image Klein (16-step distilled checkpoint) for use with ComfyUI-GGUF.

---

🎨 ComfyUI Workflow

A pre-configured Text-to-Image (T2I) workflow is included in this repository:

• Download Workflow: image_flux3_klein_distilled_t2i_GGUF.json

A pre-configured Image-edit (I2I) workflow is included in this repository:

• Download Workflow: image_flux3_klein_distilled_image_edit_GGUF.json

(Simply drag and drop the JSON file directly into ComfyUI)

---

📦 Quantization Files

Filename	Quant Type	Size	Quality Tier	Recommended Use Case / Target Hardware
flux3_image_klein_sd_Q2_K.gguf	Q2_K	1.99 GB	⚠️ Experimental	Extreme low memory / dry-run testing only (lossy, prone to visual artifacts)
flux3_image_klein_sd_Q3_K_S.gguf	Q3_K_S	2.55 GB	Acceptable	Ultra-budget / low VRAM setups (6 GB GPUs)
flux3_image_klein_sd_Q3_K_M.gguf	Q3_K_M	2.55 GB	Good	Budget 6 GB–8 GB GPUs (slightly better attention retention than _S)
flux3_image_klein_sd_Q4_K_S.gguf	Q4_K_S	3.27 GB	Very Good	8 GB GPUs (lean footprint, strong prompt adherence)
flux3_image_klein_sd_Q4_K_M.gguf	Q4_K_M	3.27 GB	Very Good	8 GB–12 GB GPUs (balanced standard for everyday inference)
flux3_image_klein_sd_Q5_K_S.gguf	Q5_K_S	3.95 GB	Excellent	10 GB–12 GB GPUs (high fidelity, lean weight footprint)
flux3_image_klein_sd_Q5_K_M.gguf	Q5_K_M	3.95 GB	⭐ Best Value	12 GB GPUs (sweet spot between visual fidelity and generation speed)
flux3_image_klein_sd_Q6_K.gguf	Q6_K

---

🖼️ Sample Generations (Q4_K_M)

---

🛠️ Setup & Usage

1. Model Placement

Place the downloaded files into your ComfyUI models directory:

📂 ComfyUI/models/
├── 📂 diffusion_models/   (or models/unet/)
│   └── flux3_image_klein_sd_Q5_K_M.gguf      <-- (Your chosen quant from this repo)
├── 📂 text_encoders/
│   └── qwen3vl_4b_int8_convrot.safetensors   <-- (From official BFL repo)
└── 📂 vae/
    └── flux2-vae.safetensors                 <-- (From official BFL repo)

Auxiliary Models Required:

• Text Encoder: qwen3vl_4b_int8_convrot.safetensors (or BF16) from BFL repo text_encoders.
• VAE: flux2-vae.safetensors from BFL repo vae.

2. Recommended Settings

• Sampler: euler
• Scheduler: simple
• Steps: 4-16
• CFG / Guidance Scale: 1.0 (Guidance is distilled into the model; leave CFG at 1.0)
• Prompting: Write descriptive, natural language prompts outlining the scene, lighting, and camera style.

---

📜 License

Governed under the FLUX Kommunity License v1.0.