NOX ENGINE

High-performance inference engine for image and video

Nox is an inference engine for NVIDIA GPUs. It specializes in executing INT8 ConvRot models as fast as possible

3× faster on average than ComfyUI running the same INT8 ConvRot checkpoints in published warm-run benchmarks.

View benchmarks

See performance

Krea 2 LTX 2.5

PERFORMANCE

Measured image and video performance on NVIDIA GPUs.

Warm-run generation time

Measured on an NVIDIA GeForce RTX 3090 · Lower is faster · seconds

Nox and ComfyUI warm-run benchmark comparison on an NVIDIA GeForce RTX 3090
Pipeline	ComfyUI	Nox	Speedup
Anima	6.15s	2.39s	2.57×
FLUX.2 Klein 4B	4.25s	1.45s	2.93×
FLUX.2 Klein 9B	7.98s	2.55s	3.13×
Ideogram 4	16.56s	8.17s	2.03×
Krea2	22.92s	7.92s	2.89×
LTX 2.3	37.52s	15.45s	2.43×
LTX 2.5	36.22s	14.98s	2.42×
MiniMax H3	41.10s	18.66s	2.20×
Qwen Image 2512	9.49s	4.00s	2.37×
Qwen Image 2.1	42.73s	13.19s	3.24×
SDXL	8.61s	6.27s	1.37×
Z-Image	12.71s	3.99s	3.19×

Both engines run the exact same INT8 ConvRot checkpoints for every model in this benchmark except SDXL, which both run in FP16.

Warm runs

Models are loaded by an initial generation. Subsequent runs use a new seed for a fresh result.

Comparison

Nox versus stock ComfyUI, both loading the exact same checkpoints. Time runs from request submission to completion, excluding output download. Speedup = ComfyUI time ÷ Nox time.

Pipeline settings and full results: Source report

Warm-run generation time

Measured on an NVIDIA GeForce RTX 4080 · Lower is faster · seconds

Nox and ComfyUI warm-run benchmark comparison on an NVIDIA GeForce RTX 4080
Pipeline	ComfyUI	Nox	Speedup
Anima	4.55s	1.53s	2.97×
FLUX.2 Klein 4B	3.65s	0.95s	3.84×
FLUX.2 Klein 9B	6.78s	1.65s	4.11×
Ideogram 4	12.91s	4.83s	2.67×
Krea2	18.53s	4.82s	3.84×
LTX 2.3	29.30s	11.17s	2.62×
LTX 2.5	28.20s	11.37s	2.48×
MiniMax H3	34.25s	11.51s	2.98×
mossimage	9.20s	2.31s	3.98×
Qwen Image 2512	9.11s	2.51s	3.63×
Qwen Image 2.1	33.91s	8.13s	4.17×
SDXL	5.98s	4.45s	1.34×
Z-Image	9.95s	2.71s	3.67×

Both engines run the exact same INT8 ConvRot checkpoints for every model in this benchmark except SDXL, which both run in FP16.

Warm runs

Models are loaded by an initial generation. Subsequent runs use a new seed for a fresh result.

Comparison

Nox versus stock ComfyUI, both loading the exact same checkpoints. Time runs from request submission to completion, excluding output download. Speedup = ComfyUI time ÷ Nox time.

Pipeline settings and full results: Source report

Warm-run generation time

Measured on an NVIDIA GeForce RTX 5090 · Lower is faster · seconds

Nox and ComfyUI warm-run benchmark comparison on an NVIDIA GeForce RTX 5090
Pipeline	ComfyUI	Nox	Speedup
Anima	3.60s	1.05s	3.43×
FLUX.2 Klein 4B	2.44s	0.72s	3.39×
FLUX.2 Klein 9B	3.58s	1.07s	3.35×
Ideogram 4	6.26s	2.92s	2.14×
Krea2	8.70s	2.73s	3.19×
LTX 2.3	19.96s	6.42s	3.11×
LTX 2.5	19.69s	6.47s	3.04×
MiniMax H3	30.63s	5.89s	5.20×
Qwen Image 2512	4.77s	1.63s	2.93×
Qwen Image 2.1	16.52s	4.80s	3.44×
SDXL	9.07s	2.27s	4.00×
Z-Image	5.85s	1.61s	3.63×

Both engines run the exact same INT8 ConvRot checkpoints for every model in this benchmark except SDXL, which both run in FP16.

Warm runs

Models are loaded by an initial generation. Subsequent runs use a new seed for a fresh result.

Comparison

Nox versus stock ComfyUI, both loading the exact same checkpoints. Time runs from request submission to completion, excluding output download. Speedup = ComfyUI time ÷ Nox time.

Pipeline settings and full results: Source report

Warm-run generation time

Measured on an NVIDIA H100 80GB HBM3 · Lower is faster · seconds

Nox and ComfyUI warm-run benchmark comparison on an NVIDIA H100 80GB HBM3
Pipeline	ComfyUI	Nox	Speedup
Anima	2.57s	0.59s	4.36×
FLUX.2 Klein 4B	1.94s	0.42s	4.62×
FLUX.2 Klein 9B	3.47s	0.63s	5.51×
Ideogram 4	7.33s	1.77s	4.14×
Krea2	10.20s	2.02s	5.05×
LTX 2.3	17.74s	5.25s	3.38×
LTX 2.5	18.09s	4.99s	3.63×
MiniMax H3	14.49s	4.05s	3.58×
Qwen Image 2512	4.47s	1.12s	3.99×
Qwen Image 2.1	18.18s	3.00s	6.06×
SDXL	3.19s	1.51s	2.11×
Z-Image	5.29s	1.05s	5.04×

Both engines run the exact same INT8 ConvRot checkpoints for every model in this benchmark except SDXL, which both run in FP16.

Warm runs

Models are loaded by an initial generation. Subsequent runs use a new seed for a fresh result.

Comparison

Nox versus stock ComfyUI, both loading the exact same checkpoints. Time runs from request submission to completion, excluding output download. Speedup = ComfyUI time ÷ Nox time.

Pipeline settings and full results: Source report

NOX ENGINE

Built for image and video inference

Specialized INT8 execution

INT8 ConvRot is a form of quantization that uses 8-bit integers for model execution and targets near-BF16 quality. Nox is built to execute these quantized models as fast as possible on NVIDIA GPUs for image and video generation.

Plan first. Execute exactly.

Nox plans execution and GPU memory use before generation, then executes that plan exactly. Available VRAM and model offloading are accounted for in advance, keeping memory use predictable across the pipeline.