Kromoss GGUF

Kromoss GGUF

GGUF quantized copies of Kromoss, a 1024 px text-to-image model fine-tuned from MOSS-Image-Base, for ComfyUI. For artists who want to generate images on a GPU with limited memory.

Made by molbal on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds GGUF quantized copies of Kromoss, a text-to-image model fine-tuned from MOSS-Image-Base at 1024 pixels. It keeps the architecture, text encoder and VAE of MOSS-Image-SFT, so it fits into existing MOSS-Image workflows. The quantized files make the model smaller to download and run.

Who it suits: It suits ComfyUI users who want to generate images from text on a graphics card with limited memory.

Industry | Who uses it | How they use it | Example
Design | Digital artists | Generate pictures from written prompts | Making concept art from a short description
Marketing | Content creators | Produce illustrations on a local machine | Creating a banner image with text placed by a layout
Software | Workflow builders | Swap in a smaller quant to fit a GPU | Choosing the smaller file on an 8 GB card

You can skip it if you do not use ComfyUI or GGUF-capable tools, since the files target that workflow.

Where to use it: It works in ComfyUI. Standard quants load with compatible GGUF tools, while the two custom quants need the comfyui-gguf-reboot loader nodes.

• ComfyUI, with the comfyui-gguf-reboot loader nodes
• Other GGUF tools, for the standard Q4_K_M, Q6_K and Q8_0 files
• Windows, macOS or Linux

Use case | Who | What they get back
Text to image | Artists | A still image from a natural-language prompt
Tag-based prompts | Anime and tag users | Images from a sequence of tags
Layout prompts | Designers | Images with text and objects placed by a JSON layout

How to set it up: Install the comfyui-gguf-reboot nodes, download one GGUF file, and place it in the diffusion_models folder. The workflow also needs a text encoder and a VAE from the upstream MOSS-Image files. The card ships a ready workflow file.

• ComfyUI
• The comfyui-gguf-reboot loader nodes
• The upstream text encoder file qwen3.5_9b_w4a8_gptq.safetensors
• The upstream VAE file flux2-vae_small.safetensors

1. Install the comfyui-gguf-reboot nodes in ComfyUI.
2. Download a Kromoss GGUF file into the diffusion_models folder.
3. Add the text encoder and the VAE to their folders.
4. Load the card's workflow file and run it.

models/diffusion_models/
models/text_encoders/
models/vae/

pip install -U huggingface_hub
hf download molbal/Kromoss-GGUF

How to use it: Load the GGUF with the UnetLoaderGGUFDynamicVRAM node, then write a prompt in plain language, as tags, or as a JSON layout. The supplied workflow uses 50 steps, CFG 3, the euler sampler and the sgm_uniform scheduler. Image sizes must be multiples of 16.

Tool | What it does
UnetLoaderGGUFDynamicVRAM | Loads a Kromoss GGUF file from the diffusion_models folder
Supplied workflow | Holds the card's reference sampling settings
Turbo LoRAs | Upstream 8-step LoRAs for faster generation

Example: You load the Q4_K_M file, type a prompt describing a scene, and run the supplied workflow at a size such as 1024 by 1024. It returns one still image that follows the prompt.

For newcomers and experts: If you are new, start with the smaller Q4_K_M file and the supplied workflow, and change only the prompt. Bigger quants trade disk space for fidelity.

If you know the field, five quants are offered. The set has three common GGML quants, plus Q8_CR and Q4_CR_L, which use special ConvRot layouts readable only through the comfyui-gguf-reboot nodes. The sampling settings include skip layer guidance on layer 22.

Written prompt, tags or JSON layout
Text encoder reads the prompt
Quantized Kromoss model denoises
VAE decodes the still image

Good to know: The Q8_CR and Q4_CR_L files do not work with general GGUF loaders.

• License: Apache 2.0, the same as the upstream model
• Cost to run: no API key, account or paid service is named, but you need a GPU and the upstream encoder and VAE files
• Data it reaches: the prompts you write and the images you create on your machine, plus files downloaded from Hugging Face
• Not for: general-purpose GGUF loaders when using the two custom quants

Title: File size in GiB
Q4_K_M: 3.35 GiB
Q4_CR_L: 3.56 GiB
Q6_K: 4.24 GiB
Q8_CR: 5.04 GiB
Q8_0: 5.31 GiB

On Hugging Face, 13 November 2026

• Task: text to image
• Parameters: 3.5B
• Weights: about 1.7 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 1,815
• Likes: 6
• License: apache-2.0
• Based on: lodestones/Kromoss
• Tags: gguf, moss-image, kromoss, comfyui, text-to-image

Links

• Model: https://huggingface.co/molbal/Kromoss-GGUF
• Files: https://huggingface.co/molbal/Kromoss-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download molbal/Kromoss-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.