Text Encoder & CLIP Models (HSWQ ConvRot INT8)
High-fidelity HSWQ ConvRot INT8 (Hybrid-Sensitivity-Weighted-Quantization) quantized weights for primary Text Encoders, CLIP, and compact language models across generative diffusion architectures (SD 1.5, SDXL, FLUX.1, SD3 / SD3.5, and Next-Gen Diffusion pipelines).
- Technical Details & Quantization Framework: Hybrid-Sensitivity-Weighted-Quantization (HSWQ)
- ComfyUI Loader for ConvRot INT8: ComfyUI-HSWQ-Loader-and-Tools
🌟 Model Overview
This repository hosts production-ready HSWQ ConvRot INT8 quantized weights for essential text encoders and conditioning backbones. By combining orthogonal Hadamard rotation ($W_{rot} = W \cdot H^T$) with layer-wise sensitivity protection, activation outlier spikes are suppressed while critical attention and normalization boundaries remain protected:
clip_l_hswq_convrot_int8.safetensors: Standalone OpenAI CLIP ViT-L/14 text encoder quantized via HSWQ ConvRot INT8 (~142.0 MB), offering massive memory savings while maintaining 100% prompt parsing integrity across SD 1.5, SDXL, and FLUX workflows.CLIP-SAE-ViT-L-14-hswq_convrot_int8.safetensors: High-fidelity ViT-L/14 text encoder with Sparse Autoencoder (SAE) feature representation, covering both the text and vision towers, with HSWQ sensitivity protection (~690.9 MB).Long-ViT-L-14-BEST-GmP-smooth-ft-full_hswq_convrot_int8.safetensors: CLIP ViT-L/14 with an extended text context (248-token position embedding), both towers, HSWQ ConvRot INT8 (~677.6 MB).ViT-L-14-BEST-smooth-GmP-HF-format_hswq_convrot_int8.safetensors: CLIP ViT-L/14 in Hugging-Face layout, both towers, HSWQ ConvRot INT8 (~678.1 MB).flan_t5_xxl_hswq_convrot_int8.safetensors: The 11B-parameter dense Google Flan-T5 XXL text encoder (encoder + decoder) quantized with HSWQ ConvRot INT8 (~14.74 GB vs ~21.49 GB FP16 baseline), reducing the footprint by ~31% to enable high-capacity prompt conditioning on consumer GPUs.Josiefied-Qwen3-0.6B-abliterated-v1_hswq_convrot_int8.safetensors: Compact 0.6B abliterated Qwen3 language model quantized with HSWQ ConvRot INT8 (~687.0 MB), providing fast, lightweight prompt expansion and text conditioning with uncensored safety vectors.
📦 Available Models
| Filename | Architecture | Upstream Base Model | Quantization | File Size | Baseline Size | VRAM Savings | Status |
|---|---|---|---|---|---|---|---|
clip_l_hswq_convrot_int8.safetensors |
OpenAI CLIP ViT-L/14 (FLUX.1 CLIP-L) | comfyanonymous/flux_text_encoders | HSWQ ConvRot INT8 | ~142.0 MB | ~234.7 MB (FP16) | ~40% Reduction | Available |
CLIP-SAE-ViT-L-14-hswq_convrot_int8.safetensors |
CLIP ViT-L/14, text + vision (SAE) | OpenAI CLIP ViT-L/14 (SAE) | HSWQ ConvRot INT8 | ~690.9 MB | ~1.63 GB (FP32) | ~58% Reduction | Available |
Long-ViT-L-14-BEST-GmP-smooth-ft-full_hswq_convrot_int8.safetensors |
CLIP ViT-L/14, extended text, text + vision | Long-ViT-L-14-BEST-GmP-smooth-ft-full |
HSWQ ConvRot INT8 | ~677.6 MB | ~1.63 GB (FP32) | ~58% Reduction | Available |
ViT-L-14-BEST-smooth-GmP-HF-format_hswq_convrot_int8.safetensors |
CLIP ViT-L/14 (HF layout), text + vision | ViT-L-14-BEST-smooth-GmP-HF-format |
HSWQ ConvRot INT8 | ~678.1 MB | ~888.3 MB (FP16) | ~24% Reduction | Available |
flan_t5_xxl_hswq_convrot_int8.safetensors |
Google Flan-T5 XXL (11B), encoder + decoder | google/flan-t5-xxl | HSWQ ConvRot INT8 | ~14.74 GB | ~21.49 GB (FP16) | ~31% Reduction | Available |
Josiefied-Qwen3-0.6B-abliterated-v1_hswq_convrot_int8.safetensors |
Qwen3 0.6B Abliterated | Josiefied-Qwen3-0.6B-abliterated-v1 | HSWQ ConvRot INT8 | ~687.0 MB | ~1.14 GB (BF16) | ~40% Reduction | Available |
🛠️ Key Features
- Orthogonal Hadamard Pre-Rotation (ConvRot): Weight matrices are pre-rotated group-wise along the input feature dimension ($W_{rot} = W \cdot H^T$) using power-of-4 normalized Hadamard blocks (
groupsize=256), eliminating catastrophic dynamic range compression caused by outlier channels. - Hybrid Sensitivity Weighting (HSWQ): Automatically measures layer sensitivity to prevent semantic collapse on critical projection and attention layers, maintaining high prompt coherence.
- Native ComfyUI Compatibility: Encodes standard
comfy_quantheader metadata ({"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256}) and per-channel scale tensors, loading automatically in ComfyUI with standardCLIPLoader,DualCLIPLoader, orTripleCLIPLoaderwithout custom nodes. - Full Precision Preservation for Critical Weights: Non-2D weights (token embeddings, position embeddings, layer normalizations, biases) are maintained in native floating-point precision to guarantee prompt parsing integrity and boundary consistency.
- Hardware Acceleration: Fully compatible with Tensor Core INT8 matrix multiplication across modern GPUs (NVIDIA Turing, Ampere, Ada Lovelace, Blackwell).
🚀 Usage in ComfyUI
Model Placement
Download the desired .safetensors files and place them into your ComfyUI text encoder directory:
ComfyUI/models/clip/
(or ComfyUI/models/LLM/ for Qwen3 language models)
Loading
Load directly via standard ComfyUI CLIP loader nodes:
- Single CLIP:
Load CLIP/CLIPLoader$\rightarrow$ selectclip_l_hswq_convrot_int8.safetensorsorCLIP-SAE-ViT-L-14-hswq_convrot_int8.safetensors. - Dual CLIP (SDXL):
DualCLIPLoader$\rightarrow$ setclip_name1toclip_l_hswq_convrot_int8.safetensors(typesdxl). - Triple CLIP (SD3 / FLUX.1):
TripleCLIPLoader$\rightarrow$ assign CLIP-L andflan_t5_xxl_hswq_convrot_int8.safetensorsaccordingly. - HSWQ Custom Node: Use ComfyUI-HSWQ-Loader-and-Tools for dedicated HSWQ graph execution and precision management.
ComfyUI automatically recognizes the comfy_quant stamp and dispatches optimized INT8 execution kernels seamlessly.
📜 Credits & License
- OpenAI CLIP ViT-L/14: openai/clip-vit-large-patch14 (MIT License)
- Google Flan-T5 XXL: google/flan-t5-xxl (Apache-2.0 License)
- Josiefied Qwen3 0.6B Abliterated: Qwen Team / Community Abliterated Releases (Apache-2.0 License)
- Quantization Pipeline: HSWQ / ConvRot INT8 Suite
- Base Licenses: Please adhere to the individual upstream licenses of the respective foundation models.