ComfyUI-GGUF-Loader Introduction
ComfyUI-GGUF-Loader is a collection of custom ComfyUI nodes for working with GGUF model files. It is the CCTech Suite fork of city96/ComfyUI-GGUF. GGUF is a model-file format popularized by llama.cpp; in this extension, its main attraction is support for quantized models that can use less GPU memory than their higher-precision counterparts.
For a familiar image-generation workflow, the simplest use is to replace ComfyUI’s Load Diffusion Model node with UNET Loader (GGUF) and select a compatible .gguf file. You can keep the rest of your workflow rather than copying a demonstration graph. The extension also offers GGUF-aware text-encoder loaders, which can reduce memory use further when a suitable quantized encoder is available.
The fork goes beyond basic image-model loading. It includes workflow nodes for audio generation, voice cloning, music, image-to-video and video-to-video work, and several model-specific image-editing tasks. These are separate tools for different workflows: you do not need to use every node to benefit from GGUF loading.
Example FLUX.1-dev GGUF image at 1024 resolution
How ComfyUI-GGUF-Loader Works
Think of quantization as packing a large set of model weights into a smaller file using less precision. This can make a model practical on a GPU with limited VRAM, although the best balance between memory use and output quality depends on the model and quantization you choose. The original project notes that transformer-based diffusion models, such as FLUX, are better candidates for this approach than older convolution-heavy UNET models.
The loader reads a compatible GGUF file and presents it to ComfyUI through the kinds of model connections your workflow expects. In a basic image workflow:
- UNET Loader (GGUF) supplies the diffusion model.
- A text-encoder loader turns your prompt into information the model can use.
- Your sampler generates the image, and a VAE decodes it for viewing or saving.
A text encoder can also take considerable memory. The extension’s CLIP Loader (GGUF) family lets you use a suitable quantized encoder in place of a regular one. These loaders can also accept supported regular safetensors or bin encoder files, so you do not have to quantize every component.
Some specialized nodes assemble more of a workflow for you. For example, an LTX audio/video models loader brings together its diffusion model, text encoder, and separate video and audio VAEs. Its preparation, sampling, and decoding nodes then handle the steps needed for that particular model family. LTX-2.3 and LTX-2.5 use different model layouts and recipes, so choose nodes that match your checkpoint.
ComfyUI-GGUF-Loader Features
GGUF model and text-encoder loading
The main loaders appear under 🤖 CCTech/GGUF:
| Node | What you use it for |
|---|---|
| UNET Loader (GGUF) | Replace Load Diffusion Model when using a compatible GGUF diffusion model. |
| UNET Loader (GGUF/Advanced) | Load the same kind of model while exposing dequant_dtype, patch_dtype, and patch_on_device controls for advanced memory and loading choices. |
| CLIP Loader (GGUF) | Load one text encoder from a supported GGUF or regular encoder file. |
| Dual / Triple / Quadruple CLIP Loader (GGUF) | Load the multiple encoders required by workflows that use two, three, or four of them. |
| Dual VAE Loader (video + audio) | Select separate video and audio VAEs in one node, with clearly named outputs. |
| Text Encoder + ClipProj Loader | Use a smaller compatible encoder with a matching projection matrix in a MiniMax-H3 workflow. |
For most artists, start with the standard loaders. The advanced UNET settings are useful when you already know what memory or compatibility issue you are trying to address. Built-in ComfyUI LoRA loaders may work with GGUF workflows, but the original README describes LoRA support here as experimental.
The Text Encoder + ClipProj Loader offers a more specific memory-saving option: for MiniMax-H3, a Qwen3-VL-4B or -8B encoder can stand in for the larger expected encoder when paired with the corresponding projection matrix. A matrix labeled mmh3-4b-* belongs with a 4B encoder; mmh3-8b-* belongs with an 8B encoder. They are not interchangeable. The author credits the projection work and matrices to ComfyUI-ClipProj and ClipProj-MiniMax-H3. This node follows ComfyUI’s normal memory management instead of pinning the encoder to a GPU.
Audio, speech, and music
- Scenema Audio, under
🤖 CCTech/Scenema, provides model loading, expressive speech generation, reference-clip voice cloning, and a separate voice-conversion node. Its generation controls include production presets, scene and language choices, pace, seed, and sentence-aware chunking for longer passages. A reference clip can guide the voice, while an optional offline voice-conversion pass aims to keep identity more consistent across chunks. The project identifies Scenema Audio models and additional components. - MiniMax Music 3, under
🤖 CCTech/MiniMax Music, pairs a models loader with an audio generator. You provide a musical caption and structured lyrics; the generator exposes controls for its generation stages and produces 44.1 kHz stereo audio. - Qwen3-TTS, under
🤖 CCTech/Qwen TTS, and OmniVoice, under🤖 CCTech/OmniVoice, provide local speech-generation workflows. The OmniVoice nodes include short-form voice cloning from a reference clip and multilingual, sentence-aware long-form speech generation. Its loader also exposes device, precision, attention, and CPU-offload choices.
Use a reference voice only when you have the appropriate rights and consent to reproduce it.
LTX audio and video workflows
The extension provides separate node groups under 🤖 CCTech/LTX-2.3 and 🤖 CCTech/LTX-2.5.
LTX-2.3 includes component loading, text-to-video, image-to-video, audio-guided video, sampling, and combined video/audio decoding. Its Video to Video (IC-LoRA) node can use a source clip as a guide for a compatible task adapter, such as a video transformation. The mode choice describes the job: t2v for a text-led start, i2v when supplying a still image, or v2v when supplying a video. The selected mode must match the connected inputs.
For IC-LoRA video guidance, ic_lora selects the task adapter and ic_lora_strength controls its influence. Choosing none lets you compare against a run without that guidance. keep_original_audio can retain the source clip’s audio as conditioning. The Crop Video Guide node removes guide material added for sampling before final decoding. More elaborate LTX-2.3 workflows in the provided README cover person removal, pose-guided replacement, and background blending; those tasks require multiple stages, not a single replacement switch.
LTX-2.5 has its own models loader, preparation nodes, sampler, guide-crop node, latent upscaler, and AV decoder. A typical image-to-video path starts with an image, samples an initial lower-resolution result, upscales its latent representation, refines it, and decodes the final video and audio. For video-to-video guidance, crop the added guide before the upscale stage. Use the frame rate from the preparation node when decoding so the resulting video and audio stay aligned.
Both LTX groups also expose reference-conditioning and face-identity tools. These use a reference image to influence the generated subject; the face-identity reinforcer is designed to pair with a specific Best-Face-ID LoRA. The provided context says that the LoRA was trained for LTX-2.3, while its pairing with LTX-2.5 is unverified.
Image guidance and editing
The provided README also describes Krea2 convenience and control nodes. A plain Krea2 img2img workflow uses a prompt and, optionally, one or more source images. Its strength affects how much the generation departs from an encoded source image. A separate Krea2 ControlNet img2img node adds control-image inputs when you want to guide structure with a compatible Control LoRA.
For that control workflow, control_mode offers several choices:
| Setting | Effect |
|---|---|
auto_depth | Make a depth map from the source image automatically. |
auto_canny | Make an edge map from the source image automatically. |
manual | Use a control image you supply, such as a pose or line-art map. |
none | Skip automatic control derivation and its requirement for a control image. |
On the Krea2 ControlNet node described in the source, an explicitly connected control_image takes precedence even when none is selected. Disconnect that input as well if you want to test a run without its control image. For pose, line art, and other manually prepared controls, see ComfyUI-ControlNet-Nodes.
Krea2 also includes an optional identity_edit mode for an appropriate Identity Edit LoRA. In this mode, the source photo becomes identity context rather than a conventional partially denoised img2img starting point, so strength no longer controls the result. The LoRA weights are separate; enabling the switch does not load them for you.
The excerpt also identifies Depth Anything V2 and Canny preprocessing used by Krea2, Qwen-Image, and Flux Klein img2img controls. Depth maps describe apparent distance in an image; Canny maps emphasize edges. They help a compatible control workflow follow composition or structure without simply reproducing the source picture. Additional node categories listed by the author include local LM Studio prompting, TripoSG image-to-3D, UniRig auto-rigging, and TRELLIS.2 / Pixal3D per-stage loaders; the supplied excerpt does not describe their controls in enough detail to give a workflow recipe here.
ComfyUI-GGUF-Loader Models
GGUF is a file format, not a single model. Choose a model that matches the kind of work you want to make, then choose an available file or quantization that suits your hardware. The provided README links these pre-quantized starting points:
| Model or component | When to consider it |
|---|---|
| FLUX.1-dev GGUF | A FLUX.1-dev image-generation workflow. |
| FLUX.1-schnell GGUF | A FLUX.1-schnell image-generation workflow. |
| Stable Diffusion 3.5 Large GGUF | A Stable Diffusion 3.5 Large workflow. |
| Stable Diffusion 3.5 Large Turbo GGUF | A workflow built for the Turbo variant. |
| T5 v1.1 XXL encoder GGUF | A compatible workflow that uses a T5 text encoder and would benefit from a quantized one. |
For specialized workflows, use the matching component set rather than mixing model families. The source identifies Scenema Audio components, MiniMax-H3 encoders and their matching projection matrices, and distinct LTX-2.3 and LTX-2.5 diffusion, text-encoder, and VAE components. It also describes Krea2 diffusion and Qwen3-VL-4B text-encoder files, plus optional task-specific LoRAs. Each of these serves a different workflow; none is a universal replacement for the others.
A quantized diffusion model may save VRAM, and a quantized text encoder may save more. That does not guarantee every combination will fit a particular GPU or produce identical results. Compare compatible options in your own workflow while keeping prompts and settings consistent.
What's New with ComfyUI-GGUF-Loader
The supplied README does not provide version numbers or a dated changelog, so these are documented additions and changes in the fork, not a version-by-version release history.
- Upstream work backported: The author lists support work for newer architectures including LTX-2, Z-Image, Ideogram-4, Qwen3-VL, and MiniMax-H3 text encoders. See PR_BACKPORT.md for the backport record.
- MiniMax-H3 loading correction: Certain buffers remain in the required precision through conversion and loading, addressing a first-sampling-step dtype crash described in the original README.
- More complete creative workflows: Scenema Audio and MiniMax Music 3 add model-loading and generation nodes, while Qwen3-TTS and OmniVoice add local speech options.
- LTX workflow nodes: Dedicated LTX-2.3 and LTX-2.5 groups bring model loading, preparation, sampling, guidance, and AV decoding into workflows tailored to each family.
- Control choices: The described img2img control nodes include a
nonemode so artists can compare controlled and uncontrolled generations more easily, subject to the behavior of any explicitly connected control image.
Troubleshooting ComfyUI-GGUF-Loader
A GGUF model is not available in the loader’s dropdown.
- Check that you are using a GGUF-aware loader, such as UNET Loader (GGUF), rather than the stock diffusion-model loader.
- Check that the file ends in
.gguf. - For a diffusion model, check the
ComfyUI/models/unetorComfyUI/models/diffusion_modelsfolder identified in the original README. For a text encoder, use the folder expected by its loader. - Reopen or refresh ComfyUI’s model selections after adding the file.
A workflow runs out of GPU memory.
- Check that you selected the intended quantized model and, if appropriate, a compatible GGUF text encoder.
- Remove any Force/Set CLIP Device node copied from an example workflow if you have only one GPU. The author states that it is not part of this node pack and is unnecessary for that setup.
- For MiniMax-H3 with ClipProj, use the matching smaller encoder and projection matrix if that workflow is appropriate for you. The pack’s loader allows ComfyUI to manage encoder paging.
An LTX checkpoint fails in an LTX loader. Check the model family first. LTX-2.5 has different components and latent geometry from LTX-2.3; loading its checkpoint through LTX-2.3 nodes is not a supported shortcut. Use the corresponding models loader, VAEs, preparation nodes, and sampling recipe.
A MiniMax-H3 ClipProj workflow gives poor or unexpected prompt guidance. Confirm that a 4B encoder uses a mmh3-4b-* matrix or an 8B encoder uses a mmh3-8b-* matrix. The <control:...> choices mentioned in the source are comparison baselines, not trained projection matrices; one deliberately ignores the prompt.
An LTX video result contains extra guide content or has audio timing problems. In an IC-LoRA workflow, include the family’s Crop Video Guide at the specified point in the chain. For LTX-2.5, crop after the first sampling pass and before latent upscale. Connect the preparation node’s frame_rate output to AV decoding rather than entering an unrelated frame rate.
A Krea2 control result does not behave as expected. Check that the Control LoRA suits the kind of map you are supplying. auto_depth creates depth guidance, auto_canny creates edge guidance, and manual uses your connected map. On the described Krea2 ControlNet node, an explicitly connected control_image overrides automatic derivation, including when none is selected.
Frequently asked questions
Do I need the example workflow? No. For basic GGUF image-model loading, the author recommends keeping your own workflow and replacing Load Diffusion Model with UNET Loader (GGUF).
Do all my model files need to be GGUF? No. The GGUF CLIP loader family accepts supported regular encoder files too, and several specialized loaders combine GGUF and other component formats.
Does choosing GGUF automatically change my image style? GGUF describes how a model is stored. Your model choice, prompt, compatible components, and workflow still determine what you generate. Quantization can introduce tradeoffs, so compare outputs when image fidelity matters.
Can I use LoRAs? The original README describes ordinary GGUF-workflow LoRA loading through ComfyUI’s built-in nodes as experimental. Some specialized workflows also require a particular task or identity LoRA and describe where it belongs in the node chain.
Learn More about ComfyUI-GGUF-Loader
- ComfyUI-GGUF-Loader project: The author’s repository and the starting point for node-specific documentation and project issues.
- Original ComfyUI-GGUF project: Background on the GGUF loader from which this fork developed.
- Backported changes: A closer look at upstream work incorporated into the fork.
- GGUF model-creation tools: Reference material for artists who later want to explore making their own quantized files.
- ComfyUI-ControlNet-Nodes: Additional pose, line-art, and other preprocessors for manually guided image workflows.
- Scenema Audio model page and ClipProj-MiniMax-H3 matrices: Component pages for those specific creative workflows.
