<h1 id="krea-2-turbo-int8-tensorwise-amd-rocm-optimized-for-comfyui">Krea 2 Turbo INT8 TensorWise — AMD ROCm Optimized for ComfyUI</h1><blockquote>⚠️ Running the models alone is not enough to unlock full AMD optimization. Follow all implementation instructions below to build a fully AMD ROCm-optimized ComfyUI stack.</blockquote>⚡ Krea 2 Turbo optimized for AMD ROCm, ComfyUI, FlashAttention, Triton/AITER, and low-memory inference.<h2 id="release-video">Release video</h2><div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/fD-J2CYetIg" start="0"></iframe></div><hr /><h1></h1>This release is the result of a full Krea 2 optimization project focused on making Krea 2 Turbo and Style Reference substantially more practical on AMD GPUs.It combines:<ul><li>Krea 2 Turbo INT8 TensorWise diffusion</li><li>Krea 2 Turbo Style Reference with the LoRA pre-fused before INT8 quantization</li><li>Krea Qwen3-VL-4B INT8 TensorWise text encoder</li><li>ROCm FlashAttention using AMD Triton/AITER</li><li>Qwen VAE Triton W8A8 acceleration</li><li>optimized Text-to-Image and Reference-to-Image ComfyUI workflows</li></ul>The diffusion checkpoints use native ComfyUI <code>int8_tensorwise</code> quantization and do not depend on ConvRot.<hr /><h2 id="performance">🚀 Performance</h2>On the same Ubuntu AMD ROCm host, using a 1368×768 / 8-step reference-to-image workflow:ConfigurationFirst workflowWarm workflowWarm samplingStock-style ComfyUI INT8 + runtime Style Reference LoRA + PyTorch cross-attention242.4s81.7s71.0sFully optimized project stack85.0s75.0s65.0s<h3 id="measured-gains">Measured gains</h3><ul><li>First workflow: 242.4s → 85.0s<ul><li>64.9% lower end-to-end time</li></ul></li><li>Warm workflow: 81.7s → 75.0s<ul><li>8.2% faster</li></ul></li><li>Warm sampling: 71.0s → 65.0s<ul><li>8.4% faster</li></ul></li><li>Iteration time: 8.87s/it → 8.16s/it<ul><li>8.0% lower</li></ul></li></ul>Cold-start results include initialization and compilation overhead. Warm sampling and sec/it are better indicators of steady-state inference performance.<h3 id="flashattention-matters">FlashAttention matters</h3>On the tested system, switching from PyTorch cross-attention to ROCm FlashAttention with AMD Triton/AITER reduced:<ul><li>warm generation: 82.5s → 75.0s</li><li>warm sampling: 71.0s → 65.0s</li><li>warm sec/it: 8.915 → 8.160</li></ul>ComfyUI's separate <code>--enable-triton-backend</code> flag produced no meaningful additional gain in this particular A/B test, so it is not required for this release.<hr /><h1 id="which-files-do-i-need">📦 Which files do I need?</h1>The full release contains three optimized model components.<h2 id="1.-krea-2-turbo-int8-tensorwise">1. Krea 2 Turbo INT8 TensorWise</h2><pre><code>krea2_turbo_int8_tensorwise_v1.safetensors</code></pre>Use this for normal Krea 2 Turbo text-to-image generation.<ul><li>all 224 main transformer GEMM weights converted to INT8 TensorWise</li><li>no ConvRot</li><li>approximately 13.16 GiB</li><li>approximately 46% smaller than the original BF16 transformer</li></ul>Place it in:<pre><code>ComfyUI/models/diffusion_models/</code></pre><hr /><h2 id="2.-krea-2-turbo-style-reference-fused-int8-tensorwise">2. Krea 2 Turbo Style Reference — Fused INT8 TensorWise</h2><pre><code>krea2_turbo_style_ref_fused_int8_tensorwise_v1.safetensors</code></pre>Use this for Krea 2 Style Reference / reference-to-image workflows.The Ostris Style Reference LoRA has already been fused into the BF16 Krea 2 Turbo model before quantization.That means:✅ No runtime Style Reference LoRA loading ✅ 224 fused transformer weights quantized to INT8 TensorWise ✅ 32 fused <code>txtfusion</code> weights retained in BF16 ✅ Reference conditioning remains available<h3 id="important">Important</h3>Do not load the Style Reference LoRA again.The LoRA is already fused into this checkpoint.You still need the Krea 2 Ostris Edit/reference-conditioning path:<a target="_blank" rel="ugc" href="https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit">https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit</a>Original Style Reference project:<a target="_blank" rel="ugc" href="https://huggingface.co/ostris/krea2_turbo_style_reference">https://huggingface.co/ostris/krea2_turbo_style_reference</a>Place the fused model in:<pre><code>ComfyUI/models/diffusion_models/</code></pre><hr /><h2 id="3.-krea-qwen3-vl-4b-int8-tensorwise-text-encoder">3. Krea Qwen3-VL-4B INT8 TensorWise Text Encoder</h2><pre><code>krea2_qwen3vl_4b_int8_tensorwise_v1.safetensors</code></pre>This converts 356 attention, MLP, and vision projection weights to native ComfyUI INT8 TensorWise.Final size is approximately 4.50 GiB, around 45% smaller than the verified BF16 source.Place it in:<pre><code>ComfyUI/models/text_encoders/</code></pre>Load it using ComfyUI <code>CLIPLoader</code> with:<pre><code>type: krea2
device: default</code></pre><hr /><h1 id="recommended-amd-rocm-setup">⚡ Recommended AMD ROCm setup</h1>For the best performance measured during this project, use:<ul><li>AMD ROCm-capable GPU</li><li>current compatible ROCm + PyTorch environment</li><li>ROCm/flash-attention</li><li>AMD Triton/AITER FlashAttention backend</li><li>Qwen VAE Triton W8A8</li><li>the optimized models from this release</li></ul>Set:<pre><code>export FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE</code></pre>Launch ComfyUI with:<pre><code>python main.py \
 --use-flash-attention \
 --disable-xformers</code></pre>Do not blindly copy ROCm, PyTorch, Triton, GFX override, or environment-variable settings from another AMD machine.GPU architecture and ROCm support should be detected first.<hr /><h1 id="copypaste-amd-rocm-installation-prompt">🧠 Copy/Paste AMD ROCm Installation Prompt</h1>I created a complete installation prompt that can be pasted into ChatGPT, Claude, Gemini, or another capable LLM with web access.It guides you interactively through:<ul><li>GPU and GFX architecture discovery</li><li>AMD driver validation</li><li>ROCm/HIP setup</li><li>Python 3.12 environment creation</li><li>compatible PyTorch + TorchVision selection</li><li>ROCm FlashAttention installation</li><li>AITER and Triton validation</li><li>ComfyUI installation</li><li>environment-variable tuning</li><li>final GPU and generation validation</li></ul>ROCm FlashAttention + AMD Triton/AITER installation prompt:<a target="_blank" rel="ugc" href="https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton#copy-and-paste-implementation-instructions-for-chatgpt-or-other-flagship-llm">https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton#copy-and-paste-implementation-instructions-for-chatgpt-or-other-flagship-llm</a><hr /><h1 id="qwen-vae-triton-w8a8">⚡ Qwen VAE Triton W8A8</h1>Krea 2's Qwen Image VAE can be a major cold-start and resolution-change bottleneck.My Qwen VAE Triton W8A8 custom node accelerates validated VAE decoder layers while keeping quality-sensitive layers in native precision.Docker isolation benchmarks showed approximately 20–21% lower VAE execution cost in the tested first-run and resolution-change cases.GitHub:<a target="_blank" rel="ugc" href="https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton">https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton</a>ComfyUI Manager search:<pre><code>Qwen VAE Triton W8A8</code></pre>Recommended preset:<pre><code>Aggressive</code></pre>Connect it as:<pre><code>VAELoader
 ↓
Patch Qwen VAE Triton W8A8
 ↓
VAEDecode</code></pre>Use the normal:<pre><code>qwen_image_vae.safetensors</code></pre>The VAE itself is not included in this model release.<hr /><h1 id="optimized-comfyui-workflows">🧩 Optimized ComfyUI Workflows</h1>Validated Text-to-Image and Style-Reference workflows are available here:<a target="_blank" rel="ugc" href="https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton/workflows">https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton/workflows</a>Full Hugging Face release:<a target="_blank" rel="ugc" href="https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton">https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton</a>The Hugging Face repository also contains:<ul><li>benchmarks</li><li>validation evidence</li><li>model provenance</li><li>tensor policies</li><li>checksums</li><li>licenses</li><li>optimized workflows</li><li>detailed AMD ROCm installation guidance</li></ul><hr /><h1 id="basic-installation">🛠️ Basic Installation</h1>Place the optimized models here:<pre><code>ComfyUI/
└── models/
 ├── diffusion_models/
 │ ├── krea2_turbo_int8_tensorwise_v1.safetensors
 │ └── krea2_turbo_style_ref_fused_int8_tensorwise_v1.safetensors
 │
 └── text_encoders/
 └── krea2_qwen3vl_4b_int8_tensorwise_v1.safetensors</code></pre>For the optimized VAE path, also install:<a target="_blank" rel="ugc" href="https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton">https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton</a>For Style Reference, install:<a target="_blank" rel="ugc" href="https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit">https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit</a><hr /><h1 id="technical-notes">🔬 Technical Notes</h1><h3 id="diffusion-quantization">Diffusion quantization</h3>The primary Krea 2 Turbo model contains 28 main transformer blocks.Eight large GEMM families per block are quantized:<pre><code>attn.wq
attn.wk
attn.wv
attn.gate
attn.wo
mlp.gate
mlp.up
mlp.down</code></pre>Total:<pre><code>28 × 8 = 224 INT8 TensorWise weights</code></pre>Each quantized layer uses native ComfyUI TensorWise metadata:<pre><code>.weight INT8
.weight_scale FP32
.comfy_quant {"format":"int8_tensorwise"}</code></pre><h3 id="style-reference-fusion">Style Reference fusion</h3>The Style Reference LoRA was fused at multiplier <code>1.0</code> while the base model was still BF16:<pre><code>W_fused = W_base + (B @ A)</code></pre>The fused transformer weights were then quantized.This avoids:<pre><code>INT8 → dequantize → merge LoRA → requantize</code></pre><hr /><h1 id="compatibility">⚠️ Compatibility</h1>This project was developed and benchmarked primarily for AMD ROCm + ComfyUI.Performance varies according to:<ul><li>GPU architecture</li><li>ROCm version</li><li>PyTorch version</li><li>FlashAttention/AITER/Triton versions</li><li>ComfyUI revision</li><li>resolution</li><li>workflow</li><li>memory configuration</li><li>kernel compilation state</li></ul>Benchmark numbers are measurements from the tested environments, not universal guarantees.<hr /><h1 id="license-and-attribution">📜 License &amp; Attribution</h1>This is an independent community derivative release.It is not an official Krea, Qwen/Alibaba, Ostris, ComfyUI, AMD, or KJnodes product, and it is not endorsed by those parties.<h3 id="krea-2-turbo-style-reference">Krea 2 Turbo / Style Reference</h3>Krea 2 Turbo and its derivatives are governed by the Krea 2 Community License Agreement.Please review the license before redistribution or commercial deployment:<a target="_blank" rel="ugc" href="https://www.krea.ai/krea-2-licensing">https://www.krea.ai/krea-2-licensing</a>Official Krea 2 Turbo:<a target="_blank" rel="ugc" href="https://huggingface.co/krea/Krea-2-Turbo">https://huggingface.co/krea/Krea-2-Turbo</a><h3 id="qwen3-vl">Qwen3-VL</h3>The upstream Qwen3-VL-4B-Instruct text encoder is published under Apache License 2.0:<a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct">https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct</a>The Hugging Face release contains the complete license copies and modification/attribution notices.<hr /><h1 id="more-projects-contact">🚀 More Projects / Contact</h1>If this project saved you VRAM, inference time, or days of debugging an AMD Krea 2 setup, check out my other optimization work.LinkedIn — Allen Barnard <a target="_blank" rel="ugc" href="https://www.linkedin.com/in/allen-b-3a35505a/">https://www.linkedin.com/in/allen-b-3a35505a/</a>YouTube — PuppetVisionAI <a target="_blank" rel="ugc" href="https://www.youtube.com/@PuppetVisionAI">https://www.youtube.com/@PuppetVisionAI</a>Website <a target="_blank" rel="ugc" href="https://puppetvision.nl">https://puppetvision.nl</a>GitHub / More Projects <a target="_blank" rel="ugc" href="https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton#work-with-me--more-projects">https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton#work-with-me--more-projects</a>I’m interested in AI Systems Engineering opportunities involving model optimization, inference systems, GPU acceleration, quantization, ROCm/CUDA, Triton, and generative-AI infrastructure.<hr /><h2 id="links">Links</h2>Full model release: <a target="_blank" rel="ugc" href="https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton">https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton</a>Optimized workflows: <a target="_blank" rel="ugc" href="https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton/workflows">https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton/workflows</a>ROCm FlashAttention / AMD Triton installation prompt: <a target="_blank" rel="ugc" href="https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton/docs/ROCM_FLASH_ATTENTION_INSTALLATION_PROMPT.md">https://huggingface.co/PuppetVision/krea-2-amd-rocm-optimized-comfy-triton/docs/ROCM_FLASH_ATTENTION_INSTALLATION_PROMPT.md</a>Qwen VAE Triton W8A8: <a target="_blank" rel="ugc" href="https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton">https://github.com/AllenCraigBarnard/ComfyUI-Qwen-VAE-Triton</a>Release video: <a target="_blank" rel="ugc" href="https://youtu.be/fD-J2CYetIg">https://youtu.be/fD-J2CYetIg</a>

Files with SHA256: 95e4f98c0e802b9cae6494b2de8407c41f32d511a9d0f28c0d85e1a840ffa180