Text Encoder & CLIP Models (HSWQ ConvRot INT8)
HSWQ ConvRot INT8 (Hybrid-Sensitivity-Weighted-Quantization) quantized weights for text encoders, CLIP, and compact language models used by SD 1.5, SDXL and FLUX.1 workflows.
The same collection also carries ConvRot W4A8 packs (asym_w4a8_int8: 4-bit weight, group size 16, ConvRot group 256), marked as such in the Quantization column.
- Technical Details & Quantization Framework: Hybrid-Sensitivity-Weighted-Quantization (HSWQ)
- ComfyUI Loader for ConvRot INT8: ComfyUI-HSWQ-Loader-and-Tools
π¦ Available Models
| Filename | Architecture | Upstream Base Model | Quantization | File Size | Baseline Size | Size Reduction | 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 | ~246.1 MB (FP16) | ~42.3% Reduction | Available |
CLIP-SAE-ViT-L-14-hswq_convrot_int8.safetensors |
CLIP ViT-L/14, text + vision (SAE) | zer0int/CLIP-SAE-ViT-L-14 | HSWQ ConvRot INT8 | ~690.9 MB | ~1.71 GB (FP32) | ~59.6% Reduction | Available |
Long-ViT-L-14-BEST-GmP-smooth-ft-full_hswq_convrot_int8.safetensors |
CLIP ViT-L/14, extended text, text + vision | zer0int/LongCLIP-GmP-ViT-L-14 | HSWQ ConvRot INT8 | ~514.4 MB | ~1.71 GB (FP32) | ~69.9% Reduction | Available |
ViT-L-14-BEST-smooth-GmP-HF-format_hswq_convrot_int8.safetensors |
CLIP ViT-L/14 (HF layout), text + vision | zer0int/CLIP-GmP-ViT-L-14 | HSWQ ConvRot INT8 | ~678.1 MB | ~931.4 MB (FP16) | ~27.2% Reduction | Available |
clip_vision_g_hswq_convrot_int8.safetensors |
CLIP ViT-bigG/14, vision | laion/CLIP-ViT-bigG-14-laion2B-39B-b160k | HSWQ ConvRot INT8 | ~1.85 GB | ~3.69 GB (FP16) | ~49.8% 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 | ~22.53 GB (FP16) | ~34.6% Reduction | Available |
Josiefied-Qwen3-0.6B-abliterated-v1_hswq_convrot_int8.safetensors |
Qwen3 0.6B Abliterated | Goekdeniz-Guelmez/Josiefied-Qwen3-0.6B-abliterated-v1 | HSWQ ConvRot INT8 | ~687.0 MB | ~1.19 GB (BF16) | ~42.4% Reduction | Available |
wevisone_vl_4b_hswq_convrot_w4a8.safetensors |
WeVisOne-VL 4B (Qwen3-VL + alpha-matte decoder), text + vision | tencent/WeVisOne-VL-4B | HSWQ ConvRot W4A8 | ~4.70 GB | ~9.70 GB (BF16) | ~51.5% Reduction | Available |
Sizes are decimal MB/GB computed from the published byte counts, and each reduction is computed from those same byte counts against the source file in the same row.
π 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$\rightarrow$ the CLIP-family packs in this repository are not applied on this route (the loading note below has the measurements), so use the float original here; the Qwen3-0.6B pack is applied on this loader (CLIPLoadertypeboogu). - Dual CLIP (SDXL):
DualCLIPLoader$\rightarrow$ theclip_lpack is not applied on this route either; setclip_name1to the floatclip_l.safetensors(typesdxl). - Triple CLIP (SD3 / FLUX.1):
TripleCLIPLoader$\rightarrow$ assign the float CLIP-L andflan_t5_xxl_hswq_convrot_int8.safetensors; the T5 pack is applied on the FLUXclip_textroute. - HSWQ Custom Node: Use ComfyUI-HSWQ-Loader-and-Tools for dedicated HSWQ graph execution and precision management.
Route-by-route measurements, including which load paths apply the INT8 and which do not: Clip-files INT8 load path.
π Security
Backdoor audit. Each file's audit status - verdict strength, weight coverage, the record behind it, and the limits of what that measurement can show - is in Backdoor Audit Verdicts. The audit code, the procedure and the raw records are in the same repository.
π Credits & License
- OpenAI CLIP ViT-L/14: openai/clip-vit-large-patch14 (MIT License)
- CLIP-SAE and CLIP-GmP checkpoints: zer0int/CLIP-SAE-ViT-L-14 and zer0int/CLIP-GmP-ViT-L-14 (MIT License each)
- LongCLIP-GmP checkpoint: zer0int/LongCLIP-GmP-ViT-L-14 (a CLIP model by OpenAI, MIT License); its declared base BeichenZhang/LongCLIP-L is Apache-2.0 License
- Google Flan-T5 XXL: google/flan-t5-xxl (Apache-2.0 License)
- Josiefied Qwen3 0.6B Abliterated: Goekdeniz-Guelmez/Josiefied-Qwen3-0.6B-abliterated-v1; its base
Qwen/Qwen3-0.6Bis Apache-2.0 License - Quantization Pipeline: HSWQ / ConvRot INT8 Suite
- Base Licenses: Please adhere to the individual upstream licenses of the respective foundation models.
Model tree for ussoewwin/Clip-files-HSWQ-ConvRot-INT8
Base model
Qwen/Qwen3-0.6B-Base