PALLER-NOVUS/MOSS-Image-GGUF

MOSS Image · NOVUS GGUF

44 GGUF artifacts for OpenMOSS MOSS-Image Base and SFT. A 5B diffusion backbone with a separate Qwen3.5-9B text encoder. Independent quantization and exploratory evaluation by NOVUS.

Download Q4_K · Download Q8_0 · Setup · Measurements · Workflows · SHA256 inventory

Q4_K is a compact starting candidate at 3.22 GB. Q8_0 is the 8-bit comparison baseline at 5.63 GB and was faster in the archived run. File size is not total RAM or VRAM usage. These measurements cover seven SFT quants, one prompt and one seed; they do not certify all 44 artifacts or every GPU.

Choose a model

Goal	Start with	What the evidence supports
Smaller download	SFT Q4_K	42.9% smaller than Q8_0; a coherent Q01 output in the recorded setup
Higher precision comparison	SFT Q8_0	Quantized 8-bit baseline, not a lossless or full-precision reference
Other quant formats	Full inventory	Availability does not establish runtime compatibility or image quality
Base model experiments	MOSS-Image-BASE-*.gguf	Separate upstream checkpoint; no Base quality benchmark is reported here

The published NOVUS file is Q4_K, not Q4_K_S. Earlier 6GB results used a differently named community artifact with no archived weight hash; they are excluded from NOVUS hardware recommendations. No 6GB/8GB guarantee or RTX 4090/5090 timing is claimed.

Get started

The GGUF contains the diffusion transformer only. You also need a compatible text encoder, VAE and runtime.

1. Download the components

Component	Download	ComfyUI folder
NOVUS SFT diffusion model	Q4_K, 3.22 GB or Q8_0, 5.63 GB	models/diffusion_models/
Community text encoder	MOSS-Qwen3.5-9B-INT8-ConvRot.safetensors	models/text_encoders/
Community VAE	MOSS-VAE.safetensors	models/vae/

The encoder and VAE are third-party components, not NOVUS weights. Their current upstream files have not been hash-matched to the historical benchmark environment.

Optional command-line download, from your ComfyUI directory:

pip install -U huggingface_hub
hf download PALLER-NOVUS/MOSS-Image-GGUF MOSS-Image-SFT-Q4_K.gguf --local-dir models/diffusion_models
hf download realrebelai/MOSS-Image_GGUFs MOSS-Qwen3.5-9B-INT8-ConvRot.safetensors --local-dir models/text_encoders
hf download realrebelai/MOSS-Image_GGUFs MOSS-VAE.safetensors --local-dir models/vae

2. Install the runtime integrations

Follow the installation instructions for ComfyUI, official OpenMOSS integration, ComfyUI-GGUF and community MOSS-GGUF bridge. Keep dependency folder names as documented by the bridge, then restart ComfyUI.

Compatibility status: the archived benchmark used local FP16 and fused-QKV runtime changes. Current upstream bridge options differ. The historical workflow records the exact graph; it does not install those changes. The BF16 example matches the current upstream precision options but has not been rerun by NOVUS. Check workflow notes before queueing. If you want the official native safetensors path instead, use OpenMOSS's ComfyUI guide; those are different model files and are not this GGUF benchmark.

3. Open the workflow and check its settings

• Historical Q4_K API graph: extracted from the actual Q01 PNG, with FP16 compute.
• Current-community BF16 API example: adapted input setting; not inference-validated.
• Download BF16 editor canvas: open this JSON in the ComfyUI editor; same untested BF16 example.
• Q01 prompt: copyable exact text used for the diagnostic image.
• Queue helper: checks node availability and enum options before submitting an API graph to your local ComfyUI.

The benchmark used 1024 × 1024, seed 42, 12 steps, Euler, sgm_uniform, CFG 3, shift 5 and skip-layer guidance on single layer 22. These describe this experiment; they are not universal optimal settings. Loading, encoder offload and decoder memory can dominate the complete pipeline.

Recorded benchmark

The chart and table come from the archived records, not GPU-tier estimates. Values below are the 1024 × 1024 observations on an RTX 3090 24GB.

SFT quant	File GB	Recorded seconds	Peak device GiB	Observed status
Q8_0	5.63	24.05	16.41	8-bit baseline
Q6_K	4.46	28.55	15.35	Generated
Q5_K	3.82	28.55	14.76	Generated
Q4_K	3.22	27.05	14.20	Compact starting candidate
IQ4_XS	3.07	31.56	14.07	Generated
Q3_K	2.58	31.56	13.60	Visible artifacts
Q2_K	2.09	27.05	13.16	Failed image

Q1_0: load failed at both tested resolutions because this runtime did not implement its dequantization. Q3_K and Q2_K produced files, but the outputs show severe degradation or a failed render. This alone does not establish an intrinsic limit of those quantization methods; runtime errors remain a possible cause.

Read the numbers correctly: one run per configuration, one seed, no repeated timing distribution. The timer spans submission to completion polling and can include queue/load/cache effects; it is not isolated denoiser time. Memory is whole-device nvidia-smi memory.used, sampled every 0.25 seconds, not process allocation or total host RAM. Archived field peak_vram_mb is preserved in JSON; the table converts those MiB readings to GiB. GB file sizes use 10⁹ bytes. An older README had different Q2_K figures; this revision uses the archived JSON: 27.05 s and 13,480 MiB.

Full methodology and limitations · CSV · JSON · Native diagnostic PNGs

What changes in the image?

Q4_K preserves a coherent scene in this example, with visible changes in face, hands, lighting and composition relative to Q8_0. This is not pixel identity or a measured zero-loss result. Q3_K shows structural artifacts; Q2_K produces an unusable image in this runtime.

The crops use fixed coordinates on the original PNGs. No retouching or selective crop movement. No BF16/F16 reference, blind reviewer scores or multi-seed quality estimate was collected for this release.

Gallery

Examples of MOSS-Image-SFT Q8_0 at 20 steps, seed 42. These demonstrate styles; they are separate from the 12-step diagnostic comparison. Exact prompts and API graphs are available in the gallery manifest.

Anime	Product

Architecture	Macro detail

Prompt-controlled scene variations are text-to-image generations. No image-input editing capability is demonstrated by this release.

NOVUS editorial artwork

The banner was generated with a separate image tool for NOVUS editorial branding. It is not MOSS output or benchmark evidence. Artwork provenance.

Sources, license and release boundaries

• Architecture and original weights: OpenMOSS, Base, SFT.
• Independent quantization and exploratory evaluation: PALLER-NOVUS. No official partnership or endorsement implied.
• Model license: Apache-2.0, as declared by the upstream model cards. License. Runtime integrations and separate pipeline components retain their own terms and attribution.
• Artifact checks: 44-file inventory with exact bytes and Hub LFS SHA256. This records publication metadata, not successful inference of every artifact. Historical loaded weight hashes, converter/source revision receipts and full environment locks are unavailable in this evidence bundle.
• Coverage: seven SFT quants generated at two resolutions; Q1_0 failed to load at both. Base and the other formats have no quality claims here. No new GPU benchmark was run for this documentation refresh.

For a useful issue report, include the exact model filename, runtime versions, GPU, available RAM/VRAM, workflow and error message. Remove credentials and private paths before sharing.