<h3 id="pure-tukano-krea2-edition">Pure Tukano – Krea2 Edition</h3>Pure Tukano expands its horizon into Krea2. Built upon the core DNA of the original SDXL release, this version is re-tuned and optimized to take full advantage of Krea2's advanced generative pipeline.If you enjoyed the aesthetic balance of the SDXL version, Pure Tukano for Krea2 takes it a step further: delivering ultra-refined skin textures, precise anatomical fidelity, highly detailed facial features, and superior lighting control. Whether you're crafting soft atmospheric portraits, vibrant lifestyle photography, or complex artistic compositions, this model provides remarkable consistency and responsiveness to your prompts.Key Features:<ul><li>Natural Texture &amp; Realism: Outstanding rendering of skin details, fine pores, and realistic light diffusion without synthetic "plastic" artifacts.</li><li>Anatomical Precision: Fine-tuned to maintain strong structural accuracy, proportion integrity, and natural posture dynamics.</li><li>Enhanced Prompt Flexibility: High versatility across various lighting conditions, camera angles, and wardrobe/aesthetic styles.</li><li>Unrestricted Expression: Retains full NSFW freedom while maintaining artistic quality, sharp focus, and high aesthetic standards.</li></ul>

Pure Tukano Krea2

INT4_CONVROT_SR (INT4 with Hadamard Weight Rotation and Stochastic Rounding) is an advanced 4-bit integer quantization format designed to run large diffusion and transformer models on consumer GPUs while minimizing the quality loss typically associated with 4-bit precision.Target GPUs (Hardware Compatibility)<ul><li>Supported Architectures: NVIDIA SM 8.0 and newer.<ul><li>RTX 3000 Series (Ampere): RTX 3060, 3070, 3080, 3090, A100.</li><li>RTX 4000 Series (Ada Lovelace): RTX 4060, 4070, 4080, 4090, H100.</li><li>RTX 5000 Series &amp; Data Center (Blackwell): Full hardware support for high-throughput INT4 Tensor Core operations.</li></ul></li><li>Note: Older architectures (GTX 10-series, RTX 20-series/Turing) either lack native INT4 Tensor Core acceleration or experience dequantization overhead that negates speed gains.</li></ul>Performance &amp; Generation Speed Benefits<ul><li>Massive VRAM Reduction: Shrinks model size by roughly ~50% compared to FP8 (and ~75% compared to FP16/BF16). This allows 12B–24B parameter models to fit entirely into 8GB–16GB VRAM without system RAM offloading/paging.</li><li>High-Throughput INT4 Tensor Cores: Leverages native integer GEMM kernels on supported GPUs, which execute operations with significantly higher theoretical throughput than floating-point math.</li><li>Memory Bandwidth Bottleneck Relief: Diffusion generation is largely memory-bandwidth bound. Streaming 4-bit weights across the GPU bus cuts the required memory traffic per sampling step in half compared to FP8, directly improving iteration speed (it/s) on cards with tighter memory buses.</li></ul>Why "CONVROT_SR"?<ul><li>ConvRot (Hadamard Rotation): Rotates weight matrices in feature space to eliminate extreme outliers before clipping, preserving high dynamic range and output visual fidelity.</li><li>SR (Stochastic Rounding): Prevents systematic quantization bias during rounding, maintaining cleaner fine details and prompt adherence compared to standard nearest-neighbor INT4 quantization.</li></ul>

Files with SHA256: a5e401fe18af5dd81884739b2ff4c7ded3f06536d509211b688f6a783e9012ac