Why Use FP32 Text Encoder? (External Site): English | Japanese
This repository hosts a fused version of the FLAN-T5-XXL model, created by combining the split files from Google's FLAN-T5-XXL repository. The files have been merged for convenience, making it easier to integrate into AI applications, including image generation workflows.
Fused for Simplicity: Combines split model files into a single, ready-to-use format.
Optimized Variants: Available in FP32, FP16, FP8, and quantized GGUF formats to balance accuracy and resource usage.
Enhanced Prompt Accuracy: Outperforms the standard T5-XXL v1.1 in generating precise outputs for image generation tasks.
Model Variant
Full
The complete T5-XXL model, including all layers (encoder + decoder). Suitable for general NLP tasks, but much larger than needed if you only want text conditioning for image
TE-only
A lightweight version containing only the text encoder portion, extracted specifically for use as a text conditioner in image generation pipelines (e.g. SD3.5, FLUX, HiDream). Much smaller than the full model since the decoder and generation-only layers are stripped out.
Model Format
FP32
Full precision, highest quality — largest file size and heaviest VRAM/compute footprint. Best for archival or when maximum fidelity matters more than speed.
FP16
Half precision with virtually no perceptible quality loss for most use cases. The recommended default — a solid balance of accuracy and file size.
INT8_ConvRot_HQ
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.
MXFP8 / FP8_svd_scaled / FP8
⚠️ 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.
GGUF (Q8_0 / Q6_K / Q5_K_M / Q4_K_M)
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.
Usage Instructions
Place the downloaded model files in one of the following directories:
models/text_encoder
models/clip
Models/CLIP
Comparison: FLAN-T5-XXL vs T5-XXL v1.1
FLAN-T5-XXL Image
T5-XXL v1.1 Image
These example images were generated using FLAN-T5-XXL and T5-XXL v1.1 models in Flux.1. FLAN-T5-XXL delivers more accurate responses to prompts.
