ComfyUI-GGUF-molbal Introduction
ComfyUI-GGUF-molbal is a fork of ComfyUI-GGUF that lets ComfyUI load supported models stored in GGUF, a format that can hold model weights at reduced precision. The fork adds support for loading Ideogram 4 and Krea 2 GGUF models, alongside other models listed below.
For an AI artist, the main benefit is access to models that might otherwise be difficult to run with limited GPU memory (VRAM). A quantized model stores its weights using fewer bits. This can reduce memory use, although smaller formats may affect image quality or performance differently depending on the model and your hardware.
The extension provides GGUF loaders for image-generation models and supported text encoders. Its usage documentation includes more guidance, pre-converted models, and sample workflows.
How ComfyUI-GGUF-molbal Works
Think of a model's weights as a large collection of numbers that guide what it creates. Quantization stores many of those numbers in a more compact form, much like packing a large set of materials into a smaller case. The model still needs to use those numbers while generating an image, so a smaller file does not automatically mean faster generation.
This approach is particularly useful for transformer-based image models, such as Flux. The original project notes that quantization was less practical for traditional convolution-heavy UNET models, while transformer and DiT models are less affected by it.
In a ComfyUI workflow, the GGUF Unet loader loads a compatible GGUF model in place of the corresponding regular model loader. The various *CLIPLoader (GGUF) nodes serve a similar purpose for supported text encoders. These text-encoder loaders can also handle regular safetensors and bin files, so you can continue using your existing CLIP model where appropriate.
ComfyUI-GGUF-molbal Features
- GGUF model loading: Use the GGUF Unet loader, found in ComfyUI's
bootlegnode category, to select a supported.ggufmodel. The original README specifiesComfyUI/models/unetas the folder for these model files. - GGUF text-encoder loading: Use a suitable
*CLIPLoader (GGUF)node where your workflow calls for a supported GGUF text encoder. You do not need to replace an existing CLIP model solely because you are using a GGUF image model. - A choice of precision levels: The supported GGUF formats range from F16 and BF16 to several 8-bit, 6-bit, 5-bit, and 4-bit options. Higher-precision choices generally favor preserving model detail; lower-bit choices prioritize compact storage. The best fit depends on your available memory, hardware, and tolerance for changes in output.
- Custom ConvRot formats: The extension includes
Q8_CR, which the author recommends for NVIDIA RTX 30-series systems. It also supportsQ4_CR(the existing W4A4 format) andQ4_CR_L, an opt-in W4A8 execution option using the same packed weight storage. These are specialized formats, not model names. - Dynamic VRAM loaders: When your ComfyUI build supports Dynamic VRAM, the author recommends its loader nodes over the legacy loaders. They avoid repeated dequantization work on large text encoders and retain native INT8 paths where available.
- K-quant acceleration: An experimental bundled Triton/CUDA backend is enabled by default for eligible K-quant linear operations. Unsupported operations fall back to a portable PyTorch path. This backend accelerates decoding but still creates a floating-point weight for the PyTorch linear operation; it is not a promise that every workflow will run faster. See the K-quant inference notes for its limits and opt-out setting.
If you are choosing among pre-converted files, the names give you a starting point. For example, Q8_0 is a general-quality 8-bit option, while Q4_0 is the smallest supported format listed in the original README. Mixed K formats such as Q4_K_M use a distribution of bit depths that protects more sensitive weights than the corresponding _S option. These are choices about how a model is stored, not different artistic styles or separate base models.
The full quantization format guide explains the available conversion options. You do not need to convert a model yourself to use an already-converted GGUF.
ComfyUI-GGUF-molbal Models
The links below lead to GGUF model collections identified in the original README. The 🍴 symbol marks collections added by this fork. These are different model families, while labels such as Q8_0 and Q4_K_M describe quantized versions of a model. Choose the model family required by your workflow, then choose an available quantization that suits your memory budget.
Image-generation and other model collections
| Model collection | When to consider it |
|---|---|
| Flux.1 Dev GGUF | When your workflow calls for Flux.1 Dev. |
| Flux.1 Schnell GGUF | When your workflow calls for Flux.1 Schnell. |
| Stable Diffusion 3.5 Large GGUF | When your workflow calls for SD 3.5 Large. |
| Stable Diffusion 3.5 Large Turbo GGUF | When your workflow calls for the Turbo variant. |
| Krea 2 GGUF 🍴 | For workflows using Krea 2; the collection includes Turbo and Raw variants. |
| Ideogram 4 GGUF 🍴 | For workflows using Ideogram 4. |
| MiniMax H3 GGUF 🍴 | For a workflow specifically built for MiniMax H3. |
| MiniMax H3 Image GGUF 🍴 | For a workflow specifically built for MiniMax H3 Image. |
| MiniMax Music3 GGUF 🍴 | For a compatible MiniMax Music3 workflow. |
| LTX 2.5 GGUF 🍴 | For a compatible LTX 2.5 workflow. |
| Qwen Image 2.1 GGUF 🍴 | For a workflow using Qwen Image 2.1. |
| Flux3 Klein GGUF 🍴 | For a workflow using Flux3 Klein. |
| Moss Image GGUF 🍴 | For a workflow using Moss Image. |
| Fibo 2.0 GGUF 🍴 | For a workflow using Fibo 2.0. |
Text-encoder collections
Text encoders help a compatible workflow interpret text prompts and, in some cases, image inputs. Select one that the workflow expects; a text encoder is not interchangeable with the image model itself.
- T5 v1.1 XXL GGUF
- Qwen3-VL 4B Instruct GGUF 🍴
- Qwen3-VL 32B Instruct GGUF 🍴
- Qwen3-VL 32B Instruct, MiniMax H3-pruned GGUF 🍴
- Qwen3.5 GGUF text encoders 🍴: The original README lists 0.8B, 2B, 4B, 9B, and 27B sizes. These require a ComfyUI build with Qwen3.5 text-encoder support. For image conditioning, place the matching
mmproj-*.ggufbeside the text encoder; text-only workflows do not need it. - Gemma 4 GGUF text encoders 🍴: The original README lists E2B, E4B, 12B, and 31B sizes and specifies ComfyUI v0.30.0 or later.
A model's name alone does not tell you which ComfyUI nodes or companion files a complete workflow needs. Start with a sample workflow or the model collection's instructions when pairing an image model with a text encoder.
What's New with ComfyUI-GGUF-molbal
The supplied README does not include a dated version history, so these are the changes it identifies rather than a version-by-version changelog:
- A maintained fork with additional model loading: The author identifies support for Ideogram 4 and Krea 2 GGUFs as a reason for this fork.
- More linked GGUF collections: The fork marks additional model collections with 🍴, including the models and text encoders listed above.
- Custom quantization choices: The README describes
Q8_CR,Q4_CR, andQ4_CR_L, as well as standard and mixed K-quant formats. These give you more ways to balance memory needs with model fidelity. - Dynamic VRAM guidance and experimental K-quant acceleration: The README recommends Dynamic VRAM loaders on supported ComfyUI builds and documents an accelerated path for certain K-quant operations.
For technical implementation details, including the custom formats and memory-mapped loading, see the author's architecture notes.
Troubleshooting ComfyUI-GGUF-molbal
The GGUF loader does not appear in ComfyUI.
- Check that you are using a compatible ComfyUI build. The original README specifies v0.27.0 or later for the custom node.
- Restart ComfyUI after adding or updating custom nodes.
- Search for GGUF Unet loader in the
bootlegnode category. For text encoders, look for a suitable*CLIPLoader (GGUF)node.
My .gguf model does not appear in the model selector.
- Check that the file has a
.ggufextension. - For the GGUF Unet loader described in the original README, check that the file is in
ComfyUI/models/unet. - Restart ComfyUI, then reopen the loader's model selector.
- Confirm that the model belongs in that loader. A text-encoder GGUF needs a compatible text-encoder loader, not the GGUF Unet loader.
My workflow cannot use a listed text encoder.
- Confirm that the workflow expects that particular encoder and size.
- For Qwen3.5, check that your ComfyUI build includes Qwen3.5 text-encoder support. If the workflow uses image conditioning, also check for the matching
mmproj-*.ggufbeside the encoder. - For Gemma 4, check that ComfyUI is v0.30.0 or later, as specified in the original README.
- Compare your workflow with the usage documentation and sample workflows.
I chose a smaller GGUF, but generation is not faster. Why?
Quantization primarily reduces how much space the model weights need. Some operations must unpack those weights while running, so speed depends on the model, selected format, loader, and hardware. On a ComfyUI build that supports it, try the recommended Dynamic VRAM loader for large text encoders.
Does a lower-bit GGUF always produce the same image as a higher-precision model?
No. Storing weights with fewer bits can change the model's output. If a compact file gives results you do not like, try a higher-precision quantization of the same model family before changing the rest of your workflow.
Can I keep using my existing CLIP file?
Yes. The original README says to use the CLIP model you were using before where appropriate; the GGUF text-encoder loaders can handle both GGUF and regular safetensors or bin files.
Where are screenshots or video walkthroughs?
None were provided in the source material. The linked documentation is the best starting point for sample workflows.
Learn More about ComfyUI-GGUF-molbal
- Usage documentation and sample workflows: Start here to see how the nodes fit into a ComfyUI workflow and find pre-converted models.
- Quantizing models: Compare the available GGUF formats when choosing between file size and precision.
- K-quant inference notes: Read this if you need to understand the experimental acceleration path or its hardware limits.
- Architecture notes: Optional deeper reading on the custom formats and memory-mapped loading.
- ComfyUI-GGUF repository: Consult the fork's source and current project documentation.
