Winnougan/Wan2.2-INT8-Convrot

Wan2.2 — INT8 ConvRot Quantized

INT8 tensor-wise quantized versions of the Wan2.2 14B diffusion models for ComfyUI, with ConvRot (group size 256) applied for maximum quantization accuracy on Ampere GPUs (RTX 30XX series).

Quantized with silveroxides/convert_to_quant directly from true BF16 source weights — single quantization pass, no double-quantization issues.

---

Download — INT8 ConvRot Models

Diffusion Models (this repo)

• wan2.2_i2v_high_noise_14B_int8_convrot.safetensors
• wan2.2_i2v_low_noise_14B_int8_convrot.safetensors

Text Encoder (this repo)

• umt5_xxl_int8_convrot.safetensors

The following files are unchanged from the official Comfy-Org release — download them separately:

LoRA

• wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors
• wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors

VAE

• wan_2.1_vae.safetensors

---

File Storage Location

ComfyUI/
├───📂 models/
│   ├───📂 diffusion_models/
│   │   ├─── wan2.2_i2v_low_noise_14B_int8_convrot.safetensors   ← this repo
│   │   └─── wan2.2_i2v_high_noise_14B_int8_convrot.safetensors  ← this repo
│   ├───📂 loras/
│   │   ├─── wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors
│   │   └─── wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors
│   ├───📂 text_encoders/
│   │   └─── umt5_xxl_int8_convrot.safetensors  ← this repo
│   └───📂 vae/
│       └─── wan_2.1_vae.safetensors

---

Quantization Details

• Format: INT8 tensor-wise (int8_tensorwise)
• ConvRot group size: 256 — all Wan2.2 dimensions (256, 4096, 5120, 13824) divide cleanly into 256, so full-strength ConvRot is applied
• Preset used: --wan (skips embeddings, encoders, head layers)
• Source: true BF16 weights — single quantization pass, no FP8→INT8 double-quantization

---

Notes

• Native INT8 support is available in recent ComfyUI builds (no extra custom nodes required for recent versions)
• For older ComfyUI builds, install BobJohnson24/ComfyUI-INT8-Fast
• These models are optimized for Ampere GPUs (RTX 30XX) — INT8 is significantly faster than FP8/FP4 on this generation

---

Disclaimer

Quantized versions of the original Wan2.2 models. All credit for the original models goes to their respective authors. Quantization may introduce minor differences in output quality compared to BF16/FP16 originals.

💬 Discord: discord.gg/CJv5wceJaN ☕ Ko-fi: ko-fi.com/winnougan