<h2 id="flan-t5-xxl-(text-encoder-only)">FLAN-T5-XXL (Text-Encoder Only)</h2><ul><li>Guide (External Site): <a target="_blank" rel="ugc" href="https://www.ai-image-journey.com/2025/03/flan-t5xxl-te-only.html">English</a> | <a target="_blank" rel="ugc" href="https://note.com/ai_image_journey/n/ncc6b1c475d8f">Japanese</a></li><li>Why Use FP32 Text Encoder? (External Site): <a target="_blank" rel="ugc" href="https://www.ai-image-journey.com/2025/08/hidream-flux1-krea.html">English</a> | <a target="_blank" rel="ugc" href="https://note.com/ai_image_journey/n/n524caae87e96">Japanese</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/easygoing0114/flan-t5-xxl-fused">Hugging Face</a></li></ul>The FP8 and GGUF format is distributed as a compressed ZIP file. Please unzip it using any decompression software of your choice before use, or download from Hugging Face page.FLAN-T5-XXL is a fine-tuned version of T5-XXL v1.1, designed to improve accuracy and performance.The original FLAN-T5-XXL model is available on <a target="_blank" rel="ugc" href="https://huggingface.co/google/flan-t5-xxl">Google's Hugging Face page</a>.When used with Flux.1, SD3.5 and HiDream, replacement for T5-XXL v1.1 to FLAN-T5-XXL offers improved prompt comprehension and enhanced image quality.This model has been streamlined by extracting only the text encoder portion, making it optimized for image generation workflows.<h2 id="model-variants">Model Variants</h2><pre><code>+------------+--------+----------+-----------+
| Flan-T5xxl | Size | Accuracy | Reccomend |
+------------+--------+----------+-----------+
| FP32 | 19 GB | 100.0 % | 🌟 |
| FP16 | 9.6 GB| 98.0 % | ✅ |
| FP8 | 4.8 GB| 95.3 % | 🔺 |
| Q8_0 | 5.1 GB| 97.6 % | ✅ |
| Q6_K | 4.0 GB| 97.3 % | 🔺 |
| Q5_K_M | 3.4 GB| 94.8 % | |
| Q4_K_M | 2.9 GB| 96.4 % | |
+------------+--------+----------+-----------+</code></pre><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a2d61c40-7c27-4593-9710-ea7d031dce89/width=525/a2d61c40-7c27-4593-9710-ea7d031dce89.jpeg" /><h2 id="usage-instructions">Usage Instructions</h2>Place the downloaded model files in one of the following directories:<ul><li><code>models/text_encoder</code></li><li><code>models/clip</code></li><li><code>Models/CLIP</code></li></ul>Select this model in place of the standard T5-XXL v1.1 model in your workflow.<h2 id="fp32-format">FP32 format</h2>The FP32 format provides the highest image quality.<h3 id="stable-diffusion-webui-forge">Stable Diffusion webUI Forge</h3>To use the text encoder in FP32 format, launch Stable Diffusion WebUI Forge with the <code>--clip-in-fp32</code> argument.<h3 id="comfyui">ComfyUI</h3>You can also use FP32 text encoders for optimal results by enabling the <code>--fp32-text-enc</code> argument at startup.<h2 id="gguf-format-in-comfyui">GGUF format in ComfyUI</h2>As of April 13, 2025, ComfyUI's default <code>DualClipLoader</code> node includes a device selection option, allowing you to choose where to load the model:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d50514d3-063e-41ed-a5ec-893047cb8546/width=525/d50514d3-063e-41ed-a5ec-893047cb8546.jpeg" /><ul><li><code>cuda</code> → VRAM</li><li><code>cpu</code> → System RAM</li></ul>Since Flux.1’s text encoder is large, setting the device to cpu and storing the model in system RAM often improves performance.Unless your system RAM is 16GB or less, keeping the model in system RAM is more effective than GGUF quantization. Thus, GGUF formats offer limited benefits in ComfyUI for most users due to sufficient RAM availability.For running Flux.1 in ComfyUI, use the FP16 or FP32 text encoder.<a target="_blank" rel="ugc" href="https://www.ai-image-journey.com/2025/03/comfyui-setting.html">More about ComfyUI settings</a>.<h2 id="comparisons">Comparisons</h2><ul><li><a target="_blank" rel="ugc" href="https://www.ai-image-journey.com/2024/12/clip-t5xxl-text-encoder.html">FLAN-T5-XXL vs T5-XXL v1.1</a></li><li><a target="_blank" rel="ugc" href="https://www.ai-image-journey.com/2024/12/image-difference-t5xxl-clip-l.html">FLAN-T5-XXL FP32 vs FP16 and Quantization</a></li></ul><h2 id="tip:-upgrade-clip-l-too">Tip: Upgrade CLIP-L Too</h2>For even better results, consider pairing FLAN-T5-XXL with an upgraded CLIP-L text encoder:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/zer0int/LongCLIP-SAE-ViT-L-14">LongCLIP-SAE-ViT-L-14</a> (ComfyUI only)</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/zer0int/CLIP-SAE-ViT-L-14">CLIP-SAE-ViT-L-14</a></li></ul>Combining FLAN-T5-XXL with an enhanced CLIP-L model can further boost image quality.<hr /><h2 id="license">License</h2><ul><li><a target="_blank" rel="ugc" href="https://www.apache.org/licenses/LICENSE-2.0)">Apache 2.0</a></li></ul>This model is based on Google's FLAN-T5-XXL, also licensed under Apache 2.0.<hr /><h2 id="update-history">Update History</h2><h3 id="august-22-2025">August 22, 2025</h3>Add Why Use FP32 Text Encoder?<h3 id="july-24-2025">July 24, 2025</h3>Re-upload of the GGUF model, reduction in model size, and correction of metadata.

FLAN-T5-XXL (Text-Encorder only)

flanT5XXLTextEncorder_mxfp8.safetensors

MXFP8

FP8_svd_scaled

Q4_K_M

FP32

FP16

Q8_0

Q6_K

Q5_K_M

Files with SHA256: 50c573425ebba20623d05da800d1f7370ebc221a242e8f6a91a82d09522f919a