MOSS-Image-SFT-GGUF 5.1B Text-to-Image Generation Model: 32GB+ VRAM

MOSS-Image-SFT-GGUF 5.1B Text-to-Image Generation Model: 32GB+ VRAM

Sample outputs are available on the model card.

At a Glance

Item Value
Repository unsloth/MOSS-Image-SFT-GGUF
Family guide MOSS-Image guide (2 articles)
Published 2026-10-22
License apache-2.0
Formats GGUF
Paper arXiv:2610.23374
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Unsloth has released “unsloth/MOSS-Image-SFT-GGUF", which is a GGUF quantized version of the 5.1B parameter text-to-image model “MOSS-Image-SFT" developed by OpenMOSS.

This model is a text-to-image model that takes natural language prompts (in English and Chinese), tag sequences, and JSON-format layout specifications as inputs to generate images. It adopts Unsloth’s proprietary “Unsloth Dynamic 2.0" technology, providing mixed-precision GGUF files where the precision of each layer is optimized based on sensitivity scans and importance matrices measured during image generation.

Specifications

The main specifications and recommended settings listed in the documentation are as follows:

  • Parameters: 5.1B (referred to as 5B in the paper)
  • Architecture: Single-stream Diffusion Transformer (DiT). Qwen3.5-9B (Qwen3.5-9B-Q8_0.gguf) is used as the text encoder, and the FLUX.2 small decoder is used for the VAE.
  • Output Specifications: Default resolution is 1024×1024 (image dimensions when using the Python API support positive multiples of 16)
  • Sampler: Flow-matching Euler
  • Recommended Steps: 50 steps
  • Recommended CFG Scale: 3.0 (Shift value is 5)
  • Layer Skipping Setting: The default guidance performs 2 transformer forward passes per step, skipping index 22 (0-indexed) of the negative branch.
  • License: Apache 2.0 license (open use including commercial use is permitted)
  • Distribution Structure Note: The GGUF files distributed this time contain only the Transformer portion. The VAE and tokenizer are to be obtained separately from unsloth/MOSS-Image-SFT-FP8, and the text encoder from unsloth/Qwen3.5-9B-GGUF to be combined.

Performance

The accuracy data presented by the publisher regarding output comparison with the bf16 version of the Transformer (measured using identical prompts, seed, initial noise, bf16 text embeddings, 1024×1024, 50 steps, CFG 3, across a total of 12 prompts consisting of 8 general prompts + 4 additional prompts) is as follows. For LPIPS, lower values indicate closer proximity to the original model, while for SSIM, values closer to 1.0 indicate higher similarity.

→ Scroll horizontally to see all columns

File Size Bits per weight LPIPS vs bf16 (mean) LPIPS vs bf16 (max) SSIM vs bf16
MOSS-Image-SFT-UD-Q3_K_XL.gguf 2.84 GB 4.47 0.281 0.493 0.713
MOSS-Image-SFT-Q4_K_M.gguf 3.20 GB 5.04 0.294 0.466 0.714
MOSS-Image-SFT-UD-Q4_K_XL.gguf 3.42 GB 5.38 0.213 0.412 0.780
MOSS-Image-SFT-UD-Q5_K_XL.gguf 3.95 GB 6.22 0.121 0.358 0.862
MOSS-Image-SFT-UD-Q6_K_XL.gguf 4.53 GB 7.13 0.106 0.354 0.883
MOSS-Image-SFT-Q8_0.gguf 5.48 GB 8.63 0.054 0.202 0.939
MOSS-Image-SFT-BF16.gguf 10.17 GB 16.00 0.000 0.000 1.000

MOSS-Image-SFT-UD-Q4_K_XL.gguf, which is recommended for general environments, has a mean LPIPS of 0.213, showing a significant improvement compared to 0.294 of the standard Q4_K_M. Furthermore, when quality closer to bf16 is desired, options include UD-Q5_K_XL (mean LPIPS 0.121), UD-Q6_K_XL (mean LPIPS 0.106), and the highest reproducibility Q8_0 (mean LPIPS 0.054, SSIM 0.939). In measurements across four evaluation prompts using stable-diffusion.cpp, Q8_0 scored 0.109, UD-Q4_K_XL scored 0.188, UD-Q3_K_XL scored 0.336, and Q4_K_M scored 0.347, indicating that the dynamic quantization (UD) versions exhibit superior reproducibility compared to standard quantization of the same bit-width.

According to the model card description, MOSS-Image is extremely sensitive to minute numerical changes, and changing even a single rounding step of bf16 in a single layer results in an average difference of 0.041 LPIPS. Therefore, it is noted that score differences around this level are closer in nature to subtle composition shifts rather than quality degradation.

Regarding the benchmark performance of the original model “MOSS-Image-SFT" itself, the technical report (arXiv:2610.23374) reports that it achieved an average text coverage rate of 96.3% on the LongLongTextBench (LLTB) benchmark evaluating long English-Chinese text rendering, recording state-of-the-art performance among open-weight models. It also reports ranking 2nd in system-level comparison and 1st in single-generator evaluation using identical enhanced prompts on RealLayoutArena (RLA), which evaluates layout-rich image generation end-to-end.

Strengths and Use Cases

The original model “MOSS-Image-SFT" excels in precise compositional control not only through natural language prompts but also via tag sequences and JSON-format layout descriptions. In particular, it boasts superior text rendering capabilities for writing text inside images, with a major feature being its ability to render long text bilingually in English and Chinese. It demonstrates strength in scenes where high information density graphics and text need to be accurately placed, such as store signs, whiteboard illustrations, or handwritten sticky notes on refrigerator doors.

Additionally, the original model is a supervised fine-tuned (SFT) variant derived from the pretrained model “MOSS-Image-Base". According to comparison examples provided by the publisher, it is adjusted to produce clearer visual grouping of text compared to the Base model, outputting structures like whiteboards with a much more organized composition. Furthermore, because it accepts JSON layout prompts in bounding box formats (yxyx, abs_centralized coordinate systems), it is well-suited for editing use cases that structurally control object placement and rendered text, as well as content production linked with layout creation tools.

The newly released GGUF quantized version, “unsloth/MOSS-Image-SFT-GGUF", aims to lower the hardware barrier required for inference while preserving the high expressive power and text rendering capabilities of the original model. Because it is subjected to mixed-precision quantization based on layer-wise importance via Unsloth Dynamic 2.0, it is suited for use cases where users want to run MOSS-Image flexibly on CPUs or Mac environments in addition to GPUs, while minimizing quality degradation.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 5.1B parameters (taken from the base model OpenMOSS-Team/MOSS-Image-SFT)

Your VRAM Quantization File size Est. memory needed
32GB (RTX 5090, etc.) BF16 25.3GB 30.4GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as GGUF.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
Qwen/Qwen-Image-2.1-Turbo 7.1B 48GB qwen-research Qwen-Image-2.1-Turbo 8-Step Fast Image Generation Model: 48GB+ VRAM (2026-11-17)
briaai/Fibo-2.0-dev 8.1B 32GB bria-fibo Fibo-2.0-dev Image Generation Model: 32GB+ VRAM, File List (2026-11-17)
Viggle/Qwen-Image-2.1-viggle-turbo 7.1B 48GB qwen-research Qwen-Image-2.1-viggle-turbo Image Generation Model: 48GB+ VRAM (2026-09-23)

Distributed Files

Weight files published in unsloth/MOSS-Image-SFT-GGUF, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
MOSS-Image-SFT-BF16.gguf 10.17GB
MOSS-Image-SFT-Q4_K_M.gguf 3.20GB
MOSS-Image-SFT-Q8_0.gguf 5.48GB
MOSS-Image-SFT-UD-Q3_K_XL.gguf 2.84GB
MOSS-Image-SFT-UD-Q4_K_XL.gguf 3.42GB
MOSS-Image-SFT-UD-Q5_K_XL.gguf 3.95GB
MOSS-Image-SFT-UD-Q6_K_XL.gguf 4.53GB

How to Get It

The weight files for “unsloth/MOSS-Image-SFT-GGUF" can be obtained from the Hugging Face repository (unsloth/MOSS-Image-SFT-GGUF). Because this model is not a gated model, anyone can download it directly without prior license agreement or access requests on Hugging Face.

The distribution format is GGUF. However, as noted above, the GGUF files provided in this repository contain only the Transformer portion. To execute image generation, you need to separately download and combine Qwen3.5-9B-Q8_0.gguf from unsloth/Qwen3.5-9B-GGUF as the text encoder, along with the set of files housed within unsloth/MOSS-Image-SFT-FP8 for the VAE and tokenizer.

Tools capable of running this model include Unsloth Studio and stable-diffusion.cpp.

Using Unsloth Studio

In Unsloth Studio, you can use the model simply by selecting “MOSS-Image-SFT" from the GUI’s “Images" tab and specifying the GGUF file from this repository. Processing is handled by the built-in stable-diffusion.cpp engine, and the required VAE, tokenizer, and Qwen3.5-9B text encoder are downloaded automatically. As standard settings, a resolution of 1024×1024, 50 steps, CFG 3, and Shift 5 are applied.

Using stable-diffusion.cpp

By using stable-diffusion.cpp with MOSS-Image support included (builds from master-975 and later), you can run it directly from the command line in CPU, Mac, and GPU environments. After gathering the required files using the hf download command, you invoke sd-cli to perform inference.

An example execution procedure is as follows:

hf download unsloth/MOSS-Image-SFT-GGUF MOSS-Image-SFT-UD-Q4_K_XL.gguf --local-dir moss
hf download unsloth/Qwen3.5-9B-GGUF Qwen3.5-9B-Q8_0.gguf --local-dir moss
hf download unsloth/MOSS-Image-SFT-FP8 vae/diffusion_pytorch_model.safetensors tokenizer/tokenizer.json --local-dir moss

./sd-cli --diffusion-model moss/MOSS-Image-SFT-UD-Q4_K_XL.gguf \
  --llm moss/Qwen3.5-9B-Q8_0.gguf \
  --vae moss/vae/diffusion_pytorch_model.safetensors \
  --tokenizer moss/tokenizer/tokenizer.json \
  --cfg-scale 3.0 --steps 50 -H 1024 -W 1024 -s 42 \
  -p "a storefront sign that says \"UNSLOTH\" in bold clean letters, daytime, sharp focus" -o moss.png

In this manner, you can easily run the latest open-weight image generation model locally through command-line tools or GUI applications in your own environment.

Related Articles

What to Read Next

Sources