MOSS-Image GGUF
This repository contains GGUF quantizations of MOSS-Image-SFT, including the supplied Turbo variant. MOSS-Image generates images from natural-language prompts, tag sequences, and JSON layouts. The SFT model and conversion are based on the OpenMOSS release, which is licensed under Apache-2.0.
Quantizations
File sizes are shown in GiB. Standard quants use common GGML formats. Proprietary quants use custom ConvRot formats supported by the comfyui-gguf-reboot loader nodes.
SFT
| Quant | Format | Size | File |
|---|---|---|---|
| Q4_0 | Standard GGUF | 3.10 GiB | moss_image_sft-Q4_0.gguf |
| Q4_K_M | Standard GGUF | 3.35 GiB | moss_image_sft-Q4_K_M.gguf |
| Q6_K | Standard GGUF | 4.24 GiB | moss_image_sft-Q6_K.gguf |
| Q8_0 | Standard GGUF | 5.31 GiB | moss_image_sft-Q8_0.gguf |
| Q8_CR | Proprietary, comfyui-gguf-reboot |
5.04 GiB | moss_image_sft-Q8_CR.gguf |
| Q4_CR_L | Proprietary, comfyui-gguf-reboot |
3.56 GiB | moss_image_sft-Q4_CR_L.gguf |
Turbo
The Turbo workflow uses 8 sampling steps.
| Quant | Format | Size | File |
|---|---|---|---|
| Q4_0 | Standard GGUF | 3.10 GiB | moss_image_sft_turbo-Q4_0.gguf |
| Q4_K_M | Standard GGUF | 3.35 GiB | moss_image_sft_turbo-Q4_K_M.gguf |
| Q6_K | Standard GGUF | 4.24 GiB | moss_image_sft_turbo-Q6_K.gguf |
| Q8_0 | Standard GGUF | 5.31 GiB | moss_image_sft_turbo-Q8_0.gguf |
| Q8_CR | Proprietary, comfyui-gguf-reboot |
5.04 GiB | moss_image_sft_turbo_Q8_CR.gguf |
| Q4_CR_L | Proprietary, comfyui-gguf-reboot |
3.56 GiB | moss_image_sft_turbo-Q4_CR_L.gguf |
Standard GGUF files can be loaded with compatible GGUF tools. Q8_CR and Q4_CR_L use custom quantization layouts and require the comfyui-gguf-reboot loader nodes. They are not compatible with general-purpose GGUF loaders.
Inputs and outputs
MOSS-Image accepts natural-language prompts, tag sequences, and JSON layouts that specify text and object placement. It generates a still image. The upstream pipeline requires image dimensions that are multiples of 16. The supplied ComfyUI workflow selects an approximately 1-megapixel portrait canvas.
The upstream SFT reference settings use 50 steps and CFG 3. The Turbo workflow included here uses 8 steps and CFG 1.
| Setting | SFT workflow | Turbo workflow |
|---|---|---|
| Steps | 50 | 8 |
| CFG | 3 | 1 |
| Sampler | euler | euler |
| Scheduler | sgm_uniform | simple |
| Shift | 5 (model default) | 5 (model default) |
| Skip layer guidance | Layer 22, full range | Bypassed, no negative branch at CFG 1 |
ComfyUI setup
Install comfyui-gguf-reboot. Use its UnetLoaderGGUFDynamicVRAM node to load a GGUF from models/diffusion_models/. These loader nodes give the best performance.
The workflow also needs the following upstream files:
| Component | File | ComfyUI folder |
|---|---|---|
| Text encoder | qwen3.5_9b_int8_convrot.safetensors | models/text_encoders/ |
| VAE | flux2-vae.safetensors | models/vae/ |
Download the 50-step SFT ComfyUI workflow.
Download the 8-step Turbo ComfyUI workflow.
License and attribution
The upstream MOSS-Image-SFT model is released under the Apache License 2.0. See the upstream model card and the OpenMOSS project for model details.
- Downloads last month
- 408
Model tree for molbal/MOSS-Image-GGUF
Base model
OpenMOSS-Team/MOSS-Image-Base