<ul><li><p>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></p></li><li><p>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></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/easygoing0114/flan-t5-xxl-fused">Hugging Face</a></p></li></ul><p></p><p><span style="color:rgb(64, 192, 87)"><strong>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.</strong></span></p><p></p><p>FLAN-T5-XXL is a fine-tuned version of T5-XXL v1.1, designed to improve accuracy and performance.</p><p></p><p>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>.</p><p></p><p>When used with Flux.1, SD3.5 and HiDream, replacement for T5-XXL v1.1 to FLAN-T5-XXL offers <strong>improved prompt comprehension</strong> and <strong>enhanced image quality</strong>.</p><p></p><p>This model has been streamlined by extracting only the <strong>text encoder</strong> portion, making it optimized for image generation workflows.</p><p></p><h2 id="model-variants">Model Variants</h2><p></p><h3 id="fp32">FP32</h3><p>Full precision, highest quality — largest file size and heaviest VRAM/compute footprint. Best for archival or when maximum fidelity matters more than speed.</p><p></p><h3 id="fp16">FP16</h3><p>Half precision with virtually no perceptible quality loss for most use cases. The recommended default — a solid balance of accuracy and file size.</p><p></p><h3 id="int8_convrot_hq">INT8_ConvRot_HQ</h3><p>Lightweight 8-bit quantization using rotation-based conversion, preserving unusually high accuracy despite the reduced size. Great if you want INT8-level savings without the usual quality trade-off.</p><p></p><h3 id="mxfp8-fp8_svd_scaled-fp8">MXFP8 / FP8_svd_scaled / FP8</h3><p>⚠️ Currently not recommended. These FP8 variants are kept for compatibility/testing but show degraded quality or stability compared to the other options — use INT8_ConvRot_HQ instead.</p><p></p><h3 id="gguf-(q8_0-q6_k-q5_k_m-q4_k_m)">GGUF (Q8_0 / Q6_K / Q5_K_M / Q4_K_M)</h3><p>Excellent compression ratios for constrained environments. Quality scales down the quantization ladder (Q8_0 closest to full precision, Q4_K_M smallest but with more quality loss) — pick based on your available memory.</p><p></p><h2 id="comparisons">Comparisons</h2><p></p><ul><li><p><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></p></li><li><p><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></p></li><li><p><a target="_blank" rel="ugc" href="https://www.ai-image-journey.com/2026/06/floating-point-precision-speed.html">INT8_ConvRot vs FP8 (Other Model)</a></p></li></ul><p></p><h2 id="tip:-upgrade-clip-l-too">Tip: Upgrade CLIP-L Too</h2><p></p><p>For even better results, consider pairing FLAN-T5-XXL with an upgraded CLIP-L text encoder:</p><p></p><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/zer0int/LongCLIP-SAE-ViT-L-14">LongCLIP-SAE-ViT-L-14</a> (ComfyUI only)</p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/zer0int/CLIP-SAE-ViT-L-14">CLIP-SAE-ViT-L-14</a></p></li></ul><p></p><p>Combining FLAN-T5-XXL with an enhanced CLIP-L model can further boost image quality.</p><p></p><hr /><h2 id="license">License</h2><ul><li><p><a target="_blank" rel="ugc" href="https://www.apache.org/licenses/LICENSE-2.0)">Apache 2.0</a></p></li></ul><p>This model is based on Google's FLAN-T5-XXL, also licensed under Apache 2.0.</p><hr /><h2 id="update-history">Update History</h2><h3 id="2026.7.10">2026.7.10</h3><ul><li><p>Upload INT8_ConvRot_HQ</p></li></ul><h3 id="2025.6.24">2025.6.24</h3><ul><li><p>Re-upload of the GGUF model, reduction in model size, and correction of metadata.</p></li></ul><p></p>

FLAN-T5-XXL (Text-Encorder only)

flanT5XXLTextEncorder_q4KM.zip

flan_t5_xxl_TE-only_Q4_K_M.zip

Q4_K_M

INT8_ConvRot_HQ

MXFP8

FP8_svd_scaled

FP32

FP16

Q8_0

Q6_K

Q5_K_M

Files with SHA256: 4240c67487e318ea60d35cbc19257138b65e7e3ff393307f22b67b7b1e292817