Team Ai
Modelpublic

martin-rizzo/Z-Image-Turbo-INT8-ConvRot-ComfyUI

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
4likes1.9kdownloads
Model Card

Z-Image-Turbo INT8-ConvRot (ComfyUI)

<img src="assets/showcase_realistic.jpg" alt="Showcase Realistic" style="display: block; margin-left: auto; margin-right: auto; width: 100%;">

This repository provides an INT8-ConvRot quantized version of the Z-Image-Turbo diffusion model. It has been specifically tailored for fast text-to-image generation within ComfyUI (or any other compatible image generation application).

The ConvRot technique applies a rotational transformation to weights and activations, effectively suppressing outliers before INT8 quantization. This process, combined with the quantization itself, enables a significant reduction in VRAM usage while maintaining excellent prompt comprehension and high-quality image generation comparable to the original BF16 weights.

Model Specifications

  • —Model Type: Scalable Single-Stream Diffusion Transformer (S3-DiT)
  • —Model Size: 6 Billion (6B) Parameters
  • —Base Model: Tongyi-MAI/Z-Image-Turbo
  • —Quantization: INT8-ConvRot (FP32 Scales) with Mixed BF16 Precision

Provided Files

  • —z_image_turbo_int8_convrot_bf16emixed.safetensors: Quantized model weights in INT8-ConvRot format with FP32 scales, preserving mixed BF16 precision in some layers.

Installation Instructions for ComfyUI

To utilize this checkpoint as a Diffusion Model within your workflow, follow these steps:

  1. 1.Download the "z_image_turbo_int8_convrot_bf16emixed.safetensors" file.
  2. 2.Place the file into the corresponding diffusion models directory of your ComfyUI installation. The path is typically: "ComfyUI/models/diffusion_models/" (or /unet depending on your version).
  3. 3.In the ComfyUI interface, load this file using a standard 'Load Diffusion Model' node.

Note: No special ConvRot-aware custom nodes are required. ComfyUI supports this type of quantization natively starting from version v0.27.0

Production Details

This quantized file was generated from the original checkpoint provided by Tongyi-MAI [[Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo)]. The conversion was performed using the z2comfy utility with the following command:

bash
./z2comfy.sh Z-Image-Turbo --int8convrot --scales-dtype FP32 --scales-trials 32 --mixed-dtype BF16E --mixed-small

Note: The `z2comfy` command is an experimental utility developed for personal use. It is not fully polished yet and may present some issues, but it is available for download at: https://github.com/martin-rizzo/ZImageCkptTools

References