中文Sign up

ethanfel/Qwen3-VL-4B-Heretic-ComfyUI-ConvRot-INT8

🤗 Hugging Face sourceapache-2.04B activated4.8 GBother✓ 1 checksumupdated today
Have the files? Submit a model torrent

If you have the complete model files and permission to share them, replace the example folder path and run this command. It checks your files, creates a torrent, and submits the magnet and checksums to Pirate Face. Keep your torrent client seeding so others can download from peers. Pirate Face does not receive the model files. Get your community key from your account. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo ethanfel/Qwen3-VL-4B-Heretic-ComfyUI-ConvRot-INT8 ./model-folder
Needs a seeder →

Qwen3-VL-4B Heretic — Dynamic ConvRot INT8 for ComfyUI Krea 2

This repository contains a true dynamic-ConvRot INT8 conversion of DreamFast/Qwen3-VL-4b-Heretic, packaged as a single-file ComfyUI text encoder for Krea 2.

Krea 2 uses Qwen3-VL-4B-Instruct conditioning, not the text-only Qwen3-4B architecture. The Heretic source is an abliterated derivative of Qwen/Qwen3-VL-4B-Instruct.

Downloads

Precision File Size Notes
INT8 qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors 4.84 GB Hosted in this repository; recommended
BF16 Upstream BF16 safetensors 8.88 GB Publicly hosted by DreamFast; not duplicated here

INT8 SHA-256:

c766f6f078bbfe7474699faaceb9484251b0c7c8dd4bf2d5730b320feca54778

Quantization

Built with silveroxides/convert_to_quant 1.3.1 at commit ee83a4d90d18f52824c3b6b23e45bc7514fb027e.

  • 356 INT8 weights with per-row FP32 scales
  • All 356 quantized layers contain convrot: true ComfyUI metadata
  • Dynamic ConvRot group sizes:
    • 216 layers at 256
    • 72 layers at 1,024
    • 68 layers at 4,096
  • 357 embeddings, norms, biases, and other sensitive tensors remain BF16
  • Deterministic seed 42
  • Prodigy learned rounding

Conversion command:

ctq \
  -i qwen3-vl-4b-heretic.safetensors \
  -o qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors \
  --int8 \
  --scaling_mode row \
  --dynamic-convrot \
  --comfy_quant \
  --save-quant-metadata \
  --low-memory \
  --device cuda \
  --heur \
  --exclude-layers \
  '(embed|norm|bias|lm_head|spiece_model|multi_modal_projector|patch_embed|patch_conv)' \
  --verbose NORMAL \
  --manual-seed 42

Fidelity versus BF16

Four varied prompts were encoded through ComfyUI's Krea 2 loader with BF16 as the reference.

Metric INT8 dynamic ConvRot
Mean cosine similarity 0.999853
Mean relative L2 error 2.13%
Mean absolute error 0.03439
Mean maximum absolute error 2.816

ComfyUI verification

GPU loading, tokenization, and end-to-end Krea 2 encoding passed with:

  • ComfyUI commit 093d571b83e7a79833200e199b46b9f5a62217f9
  • Python from the ComfyUI 13_env Conda environment
  • PyTorch 2.11.0+cu130
  • comfy-kitchen 0.2.22
  • NVIDIA GeForce RTX 5090

The output conditioning shape was (1, 10, 30720), corresponding to the 12 Qwen3-VL hidden states used by Krea 2.

Installation

Download the INT8 file into:

ComfyUI/models/text_encoders/

Select it in CLIPLoader with type krea2.

The file is a ComfyUI single-file text-encoder checkpoint, not a complete Transformers repository.

Credits