Kromoss GGUF

Kromoss is a full fine-tune of MOSS-Image-Base at 1024 px. It uses the same architecture, text encoder and VAE as MOSS-Image-SFT, so it drops into the MOSS-Image workflows. This page lists the GGUF files and the ComfyUI setup.

Hugging Face repositoryKromoss-GGUF on Hugging Face

Example output

Available Quantizations

File sizes use GiB. Standard GGUF quants are common GGML formats. Proprietary ConvRot quants require the comfyui-gguf-reboot loader.

Quant	Type	Size	File
Q4_K_M	Standard	3.35 GiB	kromoss-Q4_K_M.gguf
Q6_K	Standard	4.24 GiB	kromoss-Q6_K.gguf
Q8_0	Standard	5.31 GiB	kromoss-Q8_0.gguf
Q8_CR	Proprietary, comfyui-gguf-reboot	5.04 GiB	kromoss-Q8_CR.gguf
Q4_CR_L	Proprietary, comfyui-gguf-reboot	3.56 GiB	kromoss-Q4_CR_L.gguf

Inputs and outputs

The model accepts natural-language prompts, tag sequences, and JSON layouts that describe objects, text, and their placement. It generates still images. The upstream pipeline requires image dimensions that are multiples of 16. The supplied workflow selects a 1-megapixel portrait canvas.

ComfyUI setup

Install comfyui-gguf-reboot. Load GGUF files with its UnetLoaderGGUFDynamicVRAM node. These loader nodes give the best performance. Standard GGUF files can be opened with compatible tools, while the custom ConvRot files require this loader.

Download the text encoder and VAE from the Comfy-Org MOSS-Image repository.

Component	File	ComfyUI folder
Text encoder	qwen3.5_9b_w4a8_gptq.safetensors	models/text_encoders/
VAE	flux2-vae_small.safetensors	models/vae/

Sampling settings

The supplied workflow uses the upstream reference settings: 50 steps, CFG 3, euler sampler, sgm_uniform scheduler, and skip layer guidance on layer 22.

Setting	Value
Steps	50
CFG	3
Sampler	euler
Scheduler	sgm_uniform
Shift	5 (model default)

Workflow

Import the supplied workflow into ComfyUI.

JSON fileKromoss, 50 steps

Sources and license

The model is based on lodestones/Kromoss, which is licensed under Apache-2.0. See its model card and the OpenMOSS MOSS-Image project for model details.