Flux.2 Klein 9B Custom Ecosystem for ComfyUI

This repository provides a highly optimized, modular deployment ecosystem for the Flux.2 Klein 9B architecture. Structured specifically for advanced pipeline integration within ComfyUI, this repository contains everything required to split workloads, optimize memory layouts, and achieve top-tier generation fidelity under constrained resource environments.

---

⚠️ CRITICAL NOTICE: BASE MODEL SPECIFICATIONS

• 🚨 Full-Size Base Models: Please be aware that the primary unquantized base configurations and standard processing pipelines inside this repository have NOT been updated yet to the latest iterations.
• 💡 Current Architecture: Only the custom-optimized standalone diffusion weights, specialized Turbo/SNOFS parameters, custom text encoders, and streamlined VAE structures are fully deployed, tested, and active.

---

⚡ Optimal Settings for ComfyUI (Turbo & SNOFS Architectures)

To maximize the structural accuracy of the Flux.2 Klein 9B matrix and eliminate visual artifacts or color bleeding, we strongly recommend bypassing default sampler behaviors and adhering strictly to these optimized parameters:

🛠️ Important LoRA Tuning Mechanics:

• 🔑 The Base + Master Rule: For optimal latent stabilization, it is highly recommended to chain your selected Base Diffusion Model in conjunction with the fulx2vbase-lora-master weight layout. Running the base isolation matrix raw may bypass target structural enhancements.
• ⚠️ Rank 128 Sensitivity: When deploying the extracted High-Rank variant (flux-2-klein-9b_extracted_lora_rank_128-fp32), be aware that the high-dimensional capacity makes structural tuning significantly harder. It reacts aggressively to minor value shifts. Keep the LoRA strength scale highly constrained (try starting low at 0.3 - 0.6) to prevent sudden convergence collapse or noise corruption.

---

---

💾 Available Model Variants & Architecture

All assets are cleanly separated into dedicated subdirectories to allow dynamic VRAM loading and modular pipeline configurations:

🎭 1. Core Diffusion Weights (/diffusion_models)

These are standalone decoupled diffusion models engineered for high-performance processing, custom LoRA stacking, and split-RAM workflows.

• F29b-bf16-master-turbo_v4.safetensors (Thebest of my test 18.2 GB): Maximum fidelity master variant operating in native bfloat16 precision.
• F29b-fp8-master-turbo_V4.safetensors (9.43 GB): Optimized 8-bit precision model balancing speed and generation details.
• Flux2Klein9B-FP8-turbo.safetensors (9.08 GB): High-efficiency FP8 variant built for extreme fast-turnaround inference loops.
• Flux2Klein9B-bf16-turbo.safetensors (18.2 GB): Uncompromised precision edition designed for robust high-VRAM execution.
• Flux2Klein9B_base_bf16_SNOFS_v1.4.safetensors (18.2 GB): Fine-tuned SNOFS variant providing unique stylistic control and motion extension boundaries.
• Flux2Klein9B_3.0_fp8-turbo.safetensors (9.08 GB): Next-gen iteration compressed for low-VRAM threshold architectures.
• Flux2Klein_base-9b-kv.safetensors (18.2 GB): Specialized structural model optimized for extended key-value conditioning layouts.
• Flux2Klein_base-bf16_master_v4.safetensors (18.2 GB): Solid foundation checkpoint utilizing clean bfloat16 baseline allocations.
• Flux2Klein_master_base_v3.safetensors (18.2 GB): Stable production-grade version 3 weight set.

🗲 2. Supporting Pipeline Components

🔠 Text Encoders (/text_encoders)

Contains specialized language models optimized for advanced clip conditioning and prompt adherence:

• qwen-3-8b-flux-klein.safetensors (16.4 GB): Full-precision text encoder designed to yield maximum semantic alignment and complex prompt comprehension.
• qwen_3_8b_fp8mixed.safetensors (8.66 GB): Mixed FP8 variant engineered to optimize memory utilization and drastically cut down text processing overhead on low-VRAM environments.

🌌 Variational Autoencoders (VAE)

• /vae/fulx2vae-master.safetensors: Top-tier full precision FP32 VAE decoder, curated specifically inside the VAE directory to ensure ultimate dynamic range, color accuracy, and artifact prevention during final image decoding.
• flux.2vae-bf16.safetensors (Placed in Root): Standard ultra-lightweight bfloat16 VAE (160 MB), made available on the root boundary for rapid, memory-safe execution on constrained GPU environments.

🧩 Custom Extensions (/loras)

• check by yours.

---