Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-INT-Crush

ComfyUI Extension: ComfyUI-INT-Crush

Repo Name

ComfyUI-INT-Crush

Author
thunderfun (Account age: 3200 days)
Nodes
View all nodes(1)
Latest Updated
2026-08-23
Github Stars
0.0K

How to Install ComfyUI-INT-Crush

Install this extension via the ComfyUI Manager by searching for ComfyUI-INT-Crush
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-INT-Crush in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-INT-Crush Description

ComfyUI-INT-Crush loads quantized INT4, INT8, and OrbitQuant diffusion models in ComfyUI, with dedicated LoRA apply/remove nodes for INT4 and INT8 models and selectable inference backends.

ComfyUI-INT-Crush Introduction

ComfyUI-INT-Crush adds loader nodes for diffusion models quantized with INT-Crush Converter. Quantization stores model weights in a more compact form. Compared with fp16 weights, the author reports approximately 4× compression for INT4 and 2× compression for INT8. These are weight-size comparisons, not guarantees about total memory use, image quality, or generation speed.

You can use the extension to load compatible INT4, INT8, or OrbitQuant .safetensors files in ComfyUI. It also provides dedicated nodes for applying and removing LoRAs on INT-Crush INT4 and INT8 models.

Caution: The author states that this code has not been thoroughly tested. Keep a working copy of important workflows and check your results when trying a new model or setting. The extension was developed with AI assistance.

How ComfyUI-INT-Crush Works

A model normally stores many numerical weights that help it generate images. Think of quantization as packing those numbers into smaller containers: INT4 uses 4-bit weights, while INT8 uses 8-bit weights. During generation, the loader and its inference backend perform the calculations needed to use those packed weights.

Some INT-Crush models also use a rotation while preparing their calculations. You do not usually need to work out the rotation yourself: the loader reads the rot_size setting from compatible model metadata. Other preparation steps, such as handling padded layers, are detected automatically.

For INT4 and INT8 models, the kernel backend determines which available calculation method is used:

  • auto lets the extension select a path based on availability and the workload.
  • cuda uses the comfy-kitchen CUDA path when available.
  • triton uses a Triton-based path when available.
  • pytorch uses a PyTorch fallback. The author describes it as the slowest option, but it is useful when the faster paths are unavailable.

OrbitQuant has a different packing scheme and uses its own inference path. Changing kernel_backend does not change how OrbitQuant layers run.

ComfyUI-INT-Crush Features

INT4 and INT8 UNet loaders

Use INT4 UNet Loader (INT-Crush) for a compatible INT4 model, or INT8 UNet Loader (INT-Crush) for a compatible INT8 model. Both select a quantized .safetensors file from ComfyUI’s models/diffusion_models/ directory and provide these settings:

SettingWhat it meansPractical guidance
unet_nameThe quantized model file to load.Select the file that matches your chosen loader: INT4 or INT8.
rot_sizeThe model’s rotation size: 0, 16, 64, 256, 1024, or 4096.The loader can detect this from the file’s metadata. Use the value associated with the model rather than treating it as an image-style control.
kernel_backendThe calculation path: auto, triton, cuda, or pytorch.Start with the default, auto. Try a specific backend when diagnosing a compatibility or performance issue.

The author notes that padded layers are detected and fixed automatically. For some layers with rot_size=256, the CUDA path can combine rotation and other calculations into one step. Other layers may use separate steps. These are performance details; they are not creative controls like prompt strength.

OrbitQuant UNet loader

OrbitQuant UNet Loader (INT-Crush) selects an OrbitQuant .safetensors file from models/diffusion_models/. OrbitQuant uses a different quantization method from the INT4 and INT8 paths above. Its loader has a unet_name input but no kernel_backend setting.

INT-Crush LoRA loader and unloader

LoRA Loader (INT-Crush) applies a LoRA to the MODEL output of an INT4 or INT8 loader:

SettingWhat it means
modelThe model output from an INT-Crush INT4 or INT8 loader.
lora_nameA LoRA file from models/loras/.
strengthHow strongly to apply the LoRA; the default is 1.0.

A different strength changes how much the LoRA contributes to the result. The effect depends on the particular LoRA and model. The extension applies the LoRA without disturbing the quantized model’s rotation process.

LoRA Unloader (INT-Crush) removes INT-Crush LoRA buffers from its model input. Use it when you need a workflow branch without an attached INT-Crush LoRA.

Standard ComfyUI LoRA patching is not compatible with INT-Crush models. Use the extension’s LoRA nodes for these models instead. The readme describes the LoRA Loader’s model input specifically for INT4 and INT8 loaders; it does not establish OrbitQuant LoRA support.

ComfyUI-INT-Crush Models

The extension provides loaders for model formats, not a list of downloadable model releases. The provided documentation does not identify individual model names.

FormatWhat it offersWhen to choose it
INT4Approximately 4× smaller weights than fp16, according to the author.When you have an INT-Crush-compatible INT4 file and want the smaller of the two documented INT weight formats.
INT8Approximately 2× smaller weights than fp16, according to the author.When you have an INT-Crush-compatible INT8 file.
OrbitQuantA separate quantization scheme with its own loader and inference path.When your file is specifically an OrbitQuant .safetensors model.

Choose the loader according to how the file was quantized, not its filename alone. The source documentation does not provide image-quality comparisons between these formats, so the best choice for a particular artwork cannot be inferred from bit depth alone.

Troubleshooting ComfyUI-INT-Crush

The steps below address issues suggested by the documented inputs and compatibility notes. They are not a complete list of tested fixes.

Your model does not appear in the loader

  1. Check that the file is a .safetensors model in ComfyUI’s models/diffusion_models/ directory.
  2. Check that you are using the matching INT4, INT8, or OrbitQuant loader.
  3. Confirm that the file was produced in a format supported by the extension, such as a compatible output from INT-Crush Converter.

A backend does not work as expected

  1. Set kernel_backend to auto on an INT4 or INT8 loader.
  2. If you are investigating a faster backend, try an available alternative. pytorch is the documented fallback, though it is expected to be slower.
  3. If the model is OrbitQuant, note that the INT4/INT8 kernel_backend setting does not control its layers. OrbitQuant rotation uses Triton when available and otherwise falls back to PyTorch.

The comfy-kitchen CUDA backend does not require a Triton installation for the INT4/INT8 CUDA path. Some other paths, including RPBH and OrbitQuant rotation kernels, use Triton when it is available.

A LoRA does not behave as expected

  1. Confirm that the model input comes from an INT-Crush INT4 or INT8 loader.
  2. Select the LoRA file through lora_name from models/loras/.
  3. Use LoRA Loader (INT-Crush) rather than a standard ComfyUI LoRA patching node.
  4. Review strength; its default is 1.0. To remove an attached INT-Crush LoRA buffer, use LoRA Unloader (INT-Crush).

Frequently asked questions

Do I need to set rot_size manually?
The loader detects it from compatible safetensors metadata, including int_crush.format_version and int_crush.rot_size.

Does auto always use the same backend?
No. The documented inference path depends on the setting, available kernels, and batch size.

Will an INT4 model always generate faster than an INT8 model?
The provided documentation does not make that guarantee. It describes different calculation paths and approximate weight compression, not a universal speed comparison.

Can I use a standard ComfyUI LoRA loader?
Not for the documented INT-Crush LoRA workflow. Use the extension’s dedicated LoRA Loader and Unloader nodes.

Learn More about ComfyUI-INT-Crush

  • INT-Crush Converter explains the converter associated with compatible quantized models. It is most relevant if you want to understand where INT-Crush model files come from.
  • ComfyUI-INT-Crush repository contains the extension’s source and original documentation.

ComfyUI-INT-Crush Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

ComfyUI-INT-Crush detailed guide | ComfyUI