Key takeaways
- ComfyUI loads GGUF models through the ComfyUI-GGUF custom node, which adds GGUF loader nodes.
- GGUF files store quantised model weights, letting large image and video models run on GPUs with less memory.
- Put GGUF diffusion models in models/unet or models/diffusion_models, and GGUF text encoders in models/clip or models/text_encoders.
- Swap the standard loader for the GGUF loader in a workflow and keep the VAE and other parts the same.
- For LTX-2, Qwen image models and others, follow the model card for the matching VAE, text encoder and workflow.
To use GGUF models in ComfyUI, install the ComfyUI-GGUF custom node, put the GGUF model file in ComfyUI's unet or diffusion_models folder, restart, and replace the normal model loader in your workflow with the "Unet Loader (GGUF)" node. GGUF files are quantised, so large image and video models fit on GPUs with less memory. Quantisation is central to my FineTune Studio work with QLoRA, so this page covers what GGUF is, installing and placing files, a tutorial, examples, and LTX-2 and Qwen setups.
How do I use GGUF in ComfyUI?
Install ComfyUI-GGUF, place your .gguf diffusion model in ComfyUI/models/unet, restart ComfyUI, then load it with the "Unet Loader (GGUF)" node instead of the standard checkpoint or diffusion model loader.
Also asked as: how to gguf comfyui
Everything else in the workflow stays the same: text encoders, VAE, sampler and output nodes. If the model needs a text encoder that is also in GGUF format, load it with the GGUF CLIP loader nodes. Start from a known working workflow for the model, such as the examples in the model's repository, and swap only the loader, so you change one thing at a time when debugging.
Can I use GGUF in ComfyUI?
Yes, through the ComfyUI-GGUF custom node by city96, which adds loaders for GGUF-quantised diffusion models and text encoders. Native ComfyUI support has expanded over time, but the custom node remains the standard route for most GGUF models.
GGUF lets you run large models, such as modern image and video transformers, on consumer GPUs that cannot hold full-precision weights. The trade-off is some quality loss at lower bit widths and sometimes slower speed than optimised formats. Check that the GGUF file matches a model architecture the node supports; the repository lists supported models and updates as new ones appear.
What are GGUF models in ComfyUI?
GGUF models are model weights saved in the GGUF file format, which came from the llama.cpp project and stores tensors, often quantised to 8, 6, 5, 4 or fewer bits, together with metadata. In ComfyUI, they are usually quantised versions of diffusion and video models or text encoders.
Also asked as: what are gguf models comfyui · where to put gguf models comfyui
File names show quantisation levels, such as Q8_0, Q5_K_M or Q4_K_S. Higher numbers keep more quality and use more memory; lower numbers save memory with more quality loss. Q8 is close to full precision, Q5 and Q4 are popular balances, and lower levels suit very limited GPUs. Compare outputs on your own prompts to choose.
Where do I install the GGUF node in ComfyUI?
Install ComfyUI-GGUF into the ComfyUI/custom_nodes folder, either through ComfyUI-Manager by searching for "GGUF", or with Git, then install its Python requirement and restart ComfyUI.
Also asked as: where to install gguf comfyui
With Git:
cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF
pip install --upgrade gguf
For the Windows portable build, use the embedded Python, for example python_embeded\python.exe -s -m pip install -r ComfyUI\custom_nodes\ComfyUI-GGUF\requirements.txt from the portable folder. After restarting, search the node menu for "GGUF" to confirm the loaders appear. If they do not, check the ComfyUI console for import errors, which usually point to a missing gguf package installed into a different Python environment from the one ComfyUI uses. Keep the node updated, since support for new model architectures arrives in updates.
Where do I put GGUF files in ComfyUI?
Put GGUF diffusion or video model files in ComfyUI/models/unet or, in newer ComfyUI versions, ComfyUI/models/diffusion_models. The "Unet Loader (GGUF)" node lists files from these folders.
Also asked as: where to put gguf files comfyui
If a file does not appear in the loader's dropdown, check that it is in the right folder, has the .gguf extension, and that you restarted ComfyUI or refreshed the node list. Do not put GGUF diffusion models in the checkpoints folder, which is for full checkpoint files containing model, text encoder and VAE together. Keep a consistent naming scheme with the quantisation level in the filename.
Where do I put GGUF text encoder models in ComfyUI?
Put GGUF text encoder files, such as quantised T5 or other language encoders used by image and video models, in ComfyUI/models/clip or ComfyUI/models/text_encoders, and load them with the GGUF CLIP loader nodes.
Models such as Flux use two text encoders, so use the dual CLIP loader variant with one GGUF and one standard encoder if needed. VAE files stay in models/vae and load with the normal VAE loader. Quantising text encoders saves significant memory, because some encoders are larger than you might expect, with modest effect on prompt understanding at Q8 or Q5.
Is there a ComfyUI-GGUF tutorial?
Yes. The ComfyUI-GGUF repository's README explains installation, supported models and loader nodes, and many community video tutorials walk through specific models. A simple tutorial flow follows.
Steps:
- Install ComfyUI and confirm a standard workflow runs.
- Install ComfyUI-GGUF and its Python dependency.
- Download a GGUF quantisation of your model from its Hugging Face page.
- Place the model, text encoder and VAE in their folders.
- Open an example workflow for the model and swap in GGUF loaders.
- Generate, then compare Q8 and Q4 outputs and speed.
Monitor GPU memory to choose the highest quality that fits.
What are some ComfyUI-GGUF examples?
Common examples include running Flux image models in Q4 or Q5 on GPUs with limited memory, video models such as Wan or LTX variants with GGUF transformers, and combining GGUF models with LoRAs for styles or characters.
Model repositories on Hugging Face that publish GGUF files often include example workflows as images or JSON that you can drag into ComfyUI. Check that example workflows match your installed node versions; outdated workflows may reference renamed nodes. When sharing your own workflows, note the quantisation level and node versions so others can reproduce results.
How do I install and configure LTX-2 GGUF models in ComfyUI?
Update ComfyUI and ComfyUI-GGUF to recent versions, download an LTX-2 GGUF quantisation along with the required VAE and text encoder listed on the model card, place each in its folder, and load an official or model-card workflow, replacing the model loader with the GGUF loader.
Also asked as: how to install and configure ltx-2 gguf models in comfyui: complete 2026 guide
Video models are memory-hungry, so start with short clips, low resolution and fewer frames, then scale up. Use the sampler settings recommended on the model card, since video models are sensitive to steps and schedulers. If you run out of memory, try a lower quantisation, enable ComfyUI's memory-saving options, or offload the text encoder to the CPU. New model releases often need the latest ComfyUI version, so update before troubleshooting.
How do I use Qwen-Image-Layered GGUF in ComfyUI?
Update ComfyUI and ComfyUI-GGUF, download the GGUF quantisation and the matching text encoder and VAE from the model's page, place them in the correct folders, and start from the workflow provided by the model authors or quantisation publisher, swapping in GGUF loaders.
Also asked as: how to use qwen-image-layered gguf in comfyui: complete installation and usage guide
Layered image models output separate layers, such as foreground and background, so check the workflow's output nodes to save each layer, often as PNGs with transparency. Read the model card for supported resolutions and prompts. As with any new architecture, if loading fails, confirm your ComfyUI and node versions support the model, since support usually arrives in updates shortly after release.
Sources
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
