It does not repeat the original model card. Read NVIDIA's model card, prompt-format guidance, license, and safety notes here: nvidia/Cosmos3-Super-Text2Image.

Only transformer/ is provided as a weight artifact. The VAE, scheduler, tokenizers, safety checker, and other components are loaded from the base model.

The quantization format comes from Disty0/sdnq. SD.Next's quantization overview is here: vladmandic/sdnext Quantization.

Recipe

Setting	Value
Weights dtype	`uint4`
Static quantization	`True`
Dynamic quantization	`False`
SVD	`True`
SVD rank / steps	`32` / `8`
Quantized matmul	`False`
Dequantize FP32	`False`
Quantized conv / embedding	`False` / `False`

Quantization run: 33.74s; save time: 42.32s; transformer safetensors: 34.89 GiB.

Assemble The Pipeline

python

import json
import torch
from diffusers import Cosmos3OmniPipeline, Cosmos3OmniTransformer
from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
from huggingface_hub import snapshot_download
from sdnq.loader import load_sdnq_model

snapshot_path = snapshot_download("WaveCut/Cosmos3-Super-Text2Image-SDNQ-Uint4-SVD-Transformer")
transformer = load_sdnq_model(
    f"{snapshot_path}/transformer",
    model_cls=Cosmos3OmniTransformer,
    dtype=torch.bfloat16,
    device=torch.device("cuda"),
    dequantize_fp32=False,
    use_quantized_matmul=False,
)

pipe = Cosmos3OmniPipeline.from_pretrained(
    "nvidia/Cosmos3-Super-Text2Image",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    enable_safety_checker=True,
)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=3.0)
pipe.to("cuda")

json_caption = {
    "subjects": [],
    "background_setting": "A concise scene description.",
    "comprehensive_t2i_caption": "A detailed natural-language caption.",
    "resolution": {"H": 1024, "W": 1024},
    "aspect_ratio": "1,1",
}

result = pipe(
    prompt=json.dumps(json_caption),
    negative_prompt="",
    num_frames=1,
    height=1024,
    width=1024,
    num_inference_steps=50,
    guidance_scale=4.0,
    generator=torch.Generator(device="cuda").manual_seed(1143),
)
result.video[0].save("cosmos3_sdnq_uint4_svd.png")

load_sdnq_model expects a local path. Download this repository first, or use huggingface_hub.snapshot_download("WaveCut/Cosmos3-Super-Text2Image-SDNQ-Uint4-SVD-Transformer") and pass snapshot_path + "/transformer".

Benchmarks

Measured on one RunPod NVIDIA B200 instance with local container storage, cached model files, PyTorch 2.9.1+cu130, 1024x1024 image generation, 50 inference steps, guidance scale 4.0, flow_shift=3.0, system prompt enabled.

Transformer Component Load

Variant	Load to CUDA	VRAM after load	Torch allocated	Torch reserved	Transformer safetensors
BF16 base transformer	23.91s	122,760 MiB	122,121 MiB	122,132 MiB	119.21 GiB
SDNQ UINT4 SVD transformer	10.74s	37,096 MiB	35,760 MiB	36,376 MiB	34.89 GiB

Full Pipeline Generation

The stress set is ten handwritten JSON-caption prompts designed to stress Cyrillic text, reflections, multi-object composition, anatomy, small details, and scene-following.

Variant	Full pipeline load	VRAM after load	Torch allocated after load	Avg generation time	Min / max generation time	Peak sampled VRAM	Images
BF16 base pipeline	31.31s	125,134 MiB	124,386 MiB	16.05s	15.51s / 17.97s	141,104 MiB	10
SDNQ UINT4 SVD pipeline	23.42s	40,590 MiB	39,269 MiB	38.28s	37.74s / 40.68s	56,538 MiB	10

Original NVIDIA Example Caption

The original model repository provides assets/example_caption.json. The images below are generated locally with the same JSON-caption, seed 1143, 1024x1024, 50 steps, guidance scale 4.0.

Variant	Pipeline load	Generation time	Peak sampled VRAM
BF16 base pipeline	35.41s	18.01s	141,098 MiB
SDNQ UINT4 SVD pipeline	24.06s	74.37s	56,576 MiB

BF16 reference output:

SDNQ UINT4 SVD output:

Stress Prompt Examples

The following ten images use the same handwritten stress prompt set and seeds as the benchmark table.

Notes

This repository is an independent transformer-only quantization artifact. NVIDIA's original card states that Cosmos3-Super-Text2Image was tested in BF16; this SDNQ artifact should be treated as an experimental deployment variant and evaluated for each workload.