Krea-2_Redcraft/README.md

---
license: apache-2.0
tags:
- qwen3-vl
- text-encoder
- int8
- comfyui
- quantization
- convrot
---

# Qwen3-VL Text Encoders — INT8

INT8 quantized versions of the Qwen3-VL text encoders, for use in ComfyUI workflows (Boogu, Ideogram 4, and others that use Qwen3-VL as a text encoder).

## Files

| File | Source | ConvRot Group Size | Notes |
|---|---|---|---|
| `qwen3vl_4b_int8_convrot.safetensors` | true bf16 | 256 | Best quality, full-strength ConvRot |
| `qwen3vl_8b_int8_convrot.safetensors` | true bf16 | 16 | One intermediate dim (4304) limits group size |
| `qwen3vl_8b_int8_v2.safetensors` | true bf16 | — (no ConvRot) | Plain tensor-wise INT8 |

All quantized with [silveroxides/convert_to_quant](https://github.com/silveroxides/convert_to_quant), `--int8 --scaling_mode tensor --simple`, embedding/lm_head/norm layers excluded and kept in BF16.

## Important Note

⚠️ Do **not** use a text encoder requantized from an already-FP8-quantized source (FP8 → INT8). Double quantization causes visible color/saturation drift, especially noticeable on few-step distilled models like Boogu Turbo. These files were quantized directly from **true BF16 source weights** (`Comfy-Org/Qwen3-VL`), avoiding that issue entirely.

## Usage

Load via the standard **Load CLIP** node in ComfyUI, paired with the matching diffusion model (Boogu, Ideogram 4, etc.). Native INT8 support is available in recent ComfyUI builds; older builds need an INT8-compatible custom node.

## Credits

- Base model: [Comfy-Org/Qwen3-VL](https://huggingface.co/Comfy-Org/Qwen3-VL)
- Quantization tooling: [silveroxides/convert_to_quant](https://github.com/silveroxides/convert_to_quant)

💬 Discord: [discord.gg/CJv5wceJaN](https://discord.gg/CJv5wceJaN)
☕ Ko-fi: [ko-fi.com/winnougan](https://ko-fi.com/winnougan)