<h1>Krea 2 Turbo — QuantFunc A4W4 INT4</h1>
<p><strong>True 4-bit inference — A4W4 (4-bit activations × 4-bit weights).</strong></p>
<p>QuantFunc's core quantized matrix multiplications use <strong>INT4 activations and INT4 weights</strong> on its INT4 inference backend. A4W4 describes the compute precision used during inference, alongside the reduced weight size and memory bandwidth demand.</p>
<p>QuantFunc INT4 compresses Krea-2-Turbo to about a quarter of its 16-bit size. In our visual comparisons, composition, detail, color and style stay close to the 16-bit baseline, roughly on par with FP8 and INT8 ConvRot.</p>
<h2>Showcase</h2>
<p>All images below were generated by Krea-2-QuantFunc-4bit.</p>



Portrait photography
Sci-fi scene




<img alt="Portrait" src="https://huggingface.co/QuantFunc/Krea-2-QuantFunc-4bit/resolve/main/assets/krea2_show.webp" />
<img alt="Sci-fi" src="https://huggingface.co/QuantFunc/Krea-2-QuantFunc-4bit/resolve/main/assets/krea2_show2.webp" />


<strong>Impasto art</strong>
<strong>Watercolor illustration</strong>


<img alt="Impasto" src="https://huggingface.co/QuantFunc/Krea-2-QuantFunc-4bit/resolve/main/assets/krea2_show3.webp" />
<img alt="Watercolor" src="https://huggingface.co/QuantFunc/Krea-2-QuantFunc-4bit/resolve/main/assets/krea2_show4.webp" />



<h2>Fast denoising, fast end-to-end too</h2>
<h3>High-VRAM</h3>
<p>On an RTX 4090, same workflow and generation settings:</p>



Stage
QuantFunc INT4
FP8
Speedup




Denoising
<strong>1.6s</strong>
5.0s
<strong>3.13x</strong>


End-to-end
<strong>2.5s</strong>
6.5s
<strong>2.6x</strong>



<blockquote>
<p>End-to-end includes text encoder, denoising and VAE. Text encoder and VAE are not part of the QuantFunc plugin's acceleration path today. Actual speed varies with resolution, steps, driver, software version and hardware.</p>
</blockquote>
<h3>Low-VRAM</h3>
<p>On 8 GB / 12 GB and similar VRAM-constrained setups, 4-bit weights cut weight-bandwidth demand significantly, adding extra speedup — up to roughly <strong>11x</strong>. Actual gains depend on VRAM capacity/bandwidth, offload behavior and generation settings.</p>
<h2>Swap one loader, keep your workflow</h2>
<ol>
<li>Install or update <a href="https://github.com/QuantFunc/ComfyUI-QuantFunc" rel="ugc">ComfyUI-QuantFunc</a>.</li>
<li>Download the r128 or r32 weights.</li>
<li>Swap your model loader for the QuantFunc loader and pick the matching weight file.</li>
</ol>
<p>Every other node, connection and generation parameter stays as-is.</p>
<p>Krea-2-Turbo is a distilled turbo model — use fewer sampling steps and turn CFG off (guidance 1.0).</p>
<h2>RTX 20-series through GB300, one build covers it all</h2>
<p>Runs on every NVIDIA <strong>SM75+</strong> GPU: RTX 20/30/40/50-series, A100, H100, H200, B100, B200, GB300.</p>
<h2>Choose a model</h2>



Variant
File
Size
Use case




<strong>r128</strong>
<code>krea2-turbo-quantfunc-int4-r128.safetensors</code>
8.83 GB (8.22 GiB)
Quality-first, recommended


<strong>r32</strong>
<code>krea2-turbo-quantfunc-int4-r32.safetensors</code>
8.35 GB (7.77 GiB)
Size/VRAM-first



<h2>Loading</h2>
<p>Weights use QuantFunc's own safetensors format — load with <strong>ComfyUI-QuantFunc</strong> or the <strong>QuantFunc inference engine</strong>, not as a drop-in Diffusers checkpoint.</p>
<h2>Source &amp; license</h2>
<p>This repository distributes derived quantized weights produced from:</p>
<ul>
<li>Base model: <a href="https://huggingface.co/krea/Krea-2-Turbo" rel="ugc">krea/Krea-2-Turbo</a></li>
<li>Pretrained base: <a href="https://huggingface.co/krea/Krea-2-Raw" rel="ugc">krea/Krea-2-Raw</a></li>
</ul>
<p>Weights follow the <strong>Krea 2 Community License</strong> — please read the original model's license terms before use.</p>
<h2>Official QuantFunc links</h2>
<ul>
<li><a href="https://huggingface.co/QuantFunc" rel="ugc">Hugging Face — QuantFunc</a></li>
<li><a href="https://www.modelscope.cn/profile/QuantFunc" rel="ugc">ModelScope — QuantFunc</a></li>
<li><a href="https://discord.gg/jCp9TpFWcn" rel="ugc">Official Discord</a></li>
<li><a href="https://www.quantfunc.com/" rel="ugc">QuantFunc website</a></li>
<li><a href="https://github.com/QuantFunc/ComfyUI-QuantFunc" rel="ugc">ComfyUI-QuantFunc plugin and workflows</a></li>
</ul>
<h2>Source and weight integrity</h2>
<p>Original QuantFunc release: <a href="https://huggingface.co/QuantFunc/Krea-2-QuantFunc-4bit" rel="ugc">QuantFunc/Krea-2-QuantFunc-4bit</a>.</p>
<ul>
<li><code>krea2-turbo-quantfunc-int4-r128.safetensors</code> — SHA256: <code>cf465ba39846a5664f0d92ad3c41a8303705d91cb89cbec856371cf5635066bb</code></li>
<li><code>krea2-turbo-quantfunc-int4-r32.safetensors</code> — SHA256: <code>89cddbfaeced4fafb80b0cf284bdd29aec85ca204b5a7f5f6e245331c14cc875</code></li>
</ul>
<p>Showcase media is reproduced from the original QuantFunc model card. Exact seeds, prompts and rank variants are not supplied for every example. Performance figures are QuantFunc's reported measurements under the stated conditions; they are not guarantees for other workflows.</p>
<h2>License terms</h2><p>These quantized weights are a modified derivative of Krea 2 Turbo; they are not an official or endorsed Krea product. The Krea 2 Community License permits commercial use only below its USD 1 million company-wide annual revenue threshold; otherwise a separate Enterprise License is required. Redistribution must retain the agreement and NOTICE, and deployments must follow its content-filtering and acceptable-use requirements. See the <a href="https://cdn.jsdelivr.net/gh/krea-ai/krea-2@db3984fbc6e13b34c0064990fc2d95ac64d00058/assets/hf_samples/LICENSE.pdf" rel="ugc">full license agreement</a>.</p>

Krea 2 Turbo — QuantFunc A4W4 INT4

<p><strong>True 4-bit inference — A4W4 (4-bit activations × 4-bit weights).</strong></p>
<p>QuantFunc's core quantized matrix multiplications use <strong>INT4 activations and INT4 weights</strong> on its INT4 inference backend. A4W4 describes the compute precision used during inference, alongside the reduced weight size and memory bandwidth demand.</p>
<p>QuantFunc INT4 r128 quantization of Krea 2 Turbo. Quality-first variant. Requires ComfyUI-QuantFunc or QuantFunc. Use the Turbo model schedule, reduced sampling steps and CFG 1.0.</p><p>File: <code>krea2-turbo-quantfunc-int4-r128.safetensors</code><br />Size: 8.83 GB (8.22 GiB)<br />SHA256: <code>cf465ba39846a5664f0d92ad3c41a8303705d91cb89cbec856371cf5635066bb</code></p><p>Modified derivative of Krea 2 Turbo under the original Krea 2 Community License. Commercial use remains subject to the license revenue threshold.</p>

krea2TurboQuantfunc_a4w4INT4R128.zip

krea2-turbo-quantfunc-int4-r128.safetensors

A4W4 INT4 r128

A4W4 INT4 r32

<p><strong>True 4-bit inference — A4W4 (4-bit activations × 4-bit weights).</strong></p>
<p>QuantFunc's core quantized matrix multiplications use <strong>INT4 activations and INT4 weights</strong> on its INT4 inference backend. A4W4 describes the compute precision used during inference, alongside the reduced weight size and memory bandwidth demand.</p>
<p>QuantFunc INT4 r32 quantization of Krea 2 Turbo. Smaller rank variant. Requires ComfyUI-QuantFunc or QuantFunc. Use the Turbo model schedule, reduced sampling steps and CFG 1.0.</p><p>File: <code>krea2-turbo-quantfunc-int4-r32.safetensors</code><br />Size: 8.35 GB (7.77 GiB)<br />SHA256: <code>89cddbfaeced4fafb80b0cf284bdd29aec85ca204b5a7f5f6e245331c14cc875</code></p><p>Modified derivative of Krea 2 Turbo under the original Krea 2 Community License. Commercial use remains subject to the license revenue threshold.</p>

Files with SHA256: 542868f426bd3dfb3ece4a008c6aaa4606f5d6a06a80225d7e3610614864013e