FLUX.1 GGUF in ComfyUI: Quantization and Loader Setup Guide

To use FLUX.1 GGUF in ComfyUI, install city96's ComfyUI-GGUF nodes, load a converted checkpoint with Unet Loader (GGUF), and supply the FLUX workflow's encoders and VAE. Black Forest Labs develops the base image model, while city96 publishes the cited conversion and loader. Conversion, Loader, Workflow

This guide follows FLUX.1-dev in ComfyUI. The central question is whether a particular quantized checkpoint helps your workflow fit available resources while preserving the image details you need.

What are the key facts about FLUX.1 GGUF?

How does GGUF change a FLUX.1 workflow?

ComfyUI-GGUF adds loaders for quantized diffusion transformers and supports quantized T5 encoders. Its maintainer describes reduced bits per weight as a way to make supported image models more accessible on constrained GPUs. Node README

The conversion remains FLUX.1-dev. City96 identifies it as a direct format conversion rather than a fine-tune, which makes the original model a useful reference when evaluating a quantized copy. Conversion card

That distinction gives you a practical experiment: keep the requested scene and workflow stable while evaluating alternative representations of the same base model. Judge whether the resulting images remain useful for your deliverable.

The repository lists quantizations including Q4_0 and Q8_0. Record those labels with your checkpoint; they are not image-quality ratings or complete hardware requirements. Conversion repository

What are the limits of FLUX.1 GGUF quantization?

A GGUF diffusion-model file does not replace every component of a FLUX workflow. ComfyUI's official examples also load text encoders and a VAE, so budget for the complete graph rather than a single download. ComfyUI tutorial

ComfyUI-GGUF describes its LoRA support as experimental. Establish generation without an adapter before adding one, and test the exact adapter with the selected loader. Node README

Lower storage precision does not establish a universal speed advantage. Forge's maintainer discusses decompression overhead and competing memory-transfer costs in its GGUF explanation; evaluate the actual runtime you use. Forge explanation

The conversion card supplies no quality-retention percentage or universal GPU-memory minimum. Use a personal comparison set to decide which tradeoff is acceptable. Conversion card

How do you load FLUX.1 GGUF in ComfyUI?

Install the matching loader

Start with a working, sufficiently recent ComfyUI installation. The custom-node README requires support for custom operations when loading a diffusion model separately. Installation requirements

For a standard source installation, run the following from the ComfyUI directory with its Python environment active. This follows the repository's clone-and-dependency installation route. Node installation

git clone https://github.com/city96/ComfyUI-GGUF custom_nodes/ComfyUI-GGUF
python -m pip install --upgrade gguf

Restart ComfyUI after installation. Windows portable users should use the README's embedded-Python commands to install dependencies into the application's own environment. Portable instructions

Select one checkpoint

Choose a file from city96's FLUX.1-dev GGUF repository and record its complete filename. Put it under ComfyUI/models/unet, the directory specified by both the conversion card and the node README. Model location

Begin with one quantization choice. Downloading every available file before establishing a working loader adds uncertainty about which checkpoint produced a result and makes it harder to maintain a clear comparison.

Load an official FLUX.1-dev workflow and supply its encoders and VAE. Follow the tutorial's standalone-component arrangement when substituting the GGUF loader. Workflow, Loader

Replace the diffusion-model loader

Replace the ordinary diffusion-model loader with Unet Loader (GGUF) and select your file. Keep the remaining graph connections consistent with the original FLUX workflow. Usage instructions

Run a simple prompt before changing other parts of the graph. For example, request a folded yellow raincoat on a dark chair beside a window, with the chair fully visible. This is an original test scene, not a benchmark prompt.

Inspect whether the fabric, color, chair, and framing match the request. Record the result together with the model filename, encoder choices, VAE, dimensions, sampler settings, and application version.

If you later quantize T5, use the corresponding GGUF-capable CLIP loader described in the node README. Leave the existing CLIP model choice in place as instructed there. Text-encoder loading

Compare useful outputs

Create a small group of prompts from your real work: a close-up object, a portrait, and a composition containing lettering are useful candidates. Include the details that would cause you to reject a finished image.

For each configuration, log whether loading succeeded and how long generation took after assets were available. Keep a separate note for downloads or initial setup so they do not distort the comparison.

Review the same subjects at the same display size. Look for missing letters, altered materials, or misplaced objects. Choose acceptance criteria before deciding which image is more pleasing overall.

Avoid presenting a seed as a guarantee of identical pixels across runtimes. Use it as one recorded input while comparing observable outcomes, and document any additional workflow changes made between trials.

If a smaller file fits but produces too many rejected images, it may not be the best choice for that job. If it meets your checks, keep the working graph and outputs as a reference for future updates.

How does FLUX.1 GGUF compare with NF4 and original weights?

The ComfyUI pillar explains graph construction. For the alternative package, see the sibling FLUX NF4 guide.

What should you know before choosing FLUX.1 GGUF?

What does FLUX.1 GGUF generate?

City96's FLUX.1-dev GGUF files generate images from text descriptions through a compatible workflow. They are quantized conversions of Black Forest Labs' FLUX.1-dev model. Conversion card

Does FLUX.1 GGUF quantization reduce the parameter count?

City96's FLUX.1-dev GGUF conversion is still listed as 12B parameters. The conversion changes weight representation while retaining the base model's identity. Conversion repository

Which files do I need besides a FLUX.1 GGUF checkpoint?

A ComfyUI FLUX.1 GGUF workflow needs a GGUF-capable model loader plus the text encoders and VAE required by the FLUX graph. Use the custom-node README together with ComfyUI's official component workflow. Loader, Workflow

Which license applies to FLUX.1-dev GGUF?

City96's FLUX.1-dev GGUF conversion retains the original FLUX.1-dev Non-Commercial License and restrictions. Review the conversion's linked license before using the weights in a project. Conversion card

Sources

• FLUX.1 launch
• City96 FLUX.1-dev conversion
• ComfyUI-GGUF README
• Official ComfyUI FLUX workflow
• Forge's GGUF discussion
• NF4 package publisher
• Original FLUX.1-dev card

Related guides on PromptZone

• Realistic Photos with FLUX
• Best SDXL Models in 2026