beycanai/ControlNet-models-INT8-ConvRot

ControlNet models — INT8 ConvRot

INT8 ConvRot quantized ControlNet models for Z-Image and Qwen-Image, in ComfyUI-native .comfy_quant format. Same quality as the bf16 originals, roughly half the VRAM and disk.

Requires the loader patch: stock ComfyUI cannot load INT8 ControlNet files (the ControlNet and model-patch loaders never got native INT8 support). Install https://github.com/0xBeycan/ComfyUI-ConvRot-ControlNet — it patches the stock loader nodes, adds no new nodes, and leaves bf16 / fp8 ControlNets untouched.

Files

File	Put in	Load with	bf16 → INT8	VRAM saved	PSNR vs bf16
Z-Image-Turbo-Fun-Controlnet-Union-2.1_int8_convrot.safetensors	models/model_patches/	ModelPatchLoader	6.71 → 3.36 GB	−3.0 GB	42.0 dB
Qwen-Image-2512-Fun-Controlnet-Union-2602_int8_convrot.safetensors	models/controlnet/	Load ControlNet Model	3.51 → 1.82 GB	−1.6 GB	36.0 dB
Qwen-Image-InstantX-ControlNet-Union_int8_convrot.safetensors	models/controlnet/	Load ControlNet Model	3.54 → 1.83 GB	−1.6 GB	44.8 dB

Each file was compared against its bf16 original at the same seed and settings. Generation speed was identical on the test hardware — the gain is memory, not time. Older GPUs (30/40 series) may see a speed-up from the INT8 kernel; not tested.

Test setup: RTX 5090, ComfyUI 0.33.1, PyTorch 2.10.0+cu130.

Usage

1. Install ComfyUI-ConvRot-ControlNet into custom_nodes/ and restart ComfyUI. Confirm the console shows [ConvRot-ControlNet] patched 3/3 loaders.
2. Drop the file into the folder from the table above.
3. In your existing workflow, select the _int8_convrot file in the same loader node you already use. Nothing else changes.

Requirements

• ComfyUI ≥ 0.33 (native int8_tensorwise support in comfy.quant_ops)
• ComfyUI-ConvRot-ControlNet

How they were made

Quantized with silveroxides/convert_to_quant: INT8, ConvRot rotation (group size 256), row-wise scaling (--scaling_mode row — tensor-wise scaling breaks LoRA compatibility and softens output), .comfy_quant metadata. The zero-initialised / low-magnitude control-injection layers are kept in bf16 in every model, since quantizing them mutes the conditioning. The full recipe and per-model exclusion lists are in the loader repo's README.

Sources

• alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1
• alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union (2602 release)
• InstantX/Qwen-Image-ControlNet-Union

All three originals are Apache-2.0; these quantized files carry the same license.