Home
Services
Local AI Deployment Test Results Image & Video NeoCloud Results Offline Private Security System Digital Transformation Professional Photography
Team
Bruce Experience Education Contact

Image and Video Generation Results

Diffusion models for images and video, measured in ComfyUI on our dual Radeon RX 7900 XTX workstation and our RTX 3080 Ti + RTX 3080 workstation. The workflows and raw results are published on GitHub.Image and video generation times from our AMD and NVIDIA workstations.

AMD Image and Video Generation

Diffusion models in ComfyUI on our AMD workstation. Since November 7, MiniMax H3 INT8 text to video runs across both cards with tensor parallelism over direct GPU-to-GPU transfers; the other workflows keep the denoiser on one card and the text encoder and VAE on the other.Image and video models in ComfyUI, split across both cards.

One complete request each, from submitting the job to the saved output, as recorded by our test runner on November 7, 2026, in the final card layout (rows marked Nov 5 run: the earlier November 5 run). ComfyUI 0.44 on ROCm 7.2, both cards in use; results vary with the model, settings and hardware. MiniMax H3 Image: November 15, 2026, on ComfyUI 0.45.4, with its files already cached in RAM.

Samples under each model were generated on the same workstation with the same workflows, during our November 5, 2026 benchmark runs.

Qwen-Image 2.1 BF16

Model / precisionTaskOutputStepsTime
Qwen-Image 2.1 BF16Text to image512×5124090.3 s
Qwen-Image 2.1 BF16Image edit640×4164025.94 sNov 5 run

FLUX 3 Image Klein BF16

Model / precisionTaskOutputStepsTime
FLUX 3 Image Klein BF16Text to image1024×10241696.5 s
FLUX 3 Image Klein BF16Image edit1248×8321642.1 s

MiniMax H3 Image INT8

Model / precisionTaskOutputStepsTime
MiniMax H3 Image INT8Text to image2048×2048636.3 s
MiniMax H3 Image INT8Text to image4096×40966126.3 s
MiniMax H3 Image INT8Image edit2048×2048664.9 s

MiniMax H3 INT8

Model / precisionTaskOutputStepsTime
MiniMax H3 INT8Video with audio608×352, 39 frames2056.4 smodel already loaded
MiniMax H3 INT8Video with audio960×544, 39 frames20132.3 s

MiniMax H3 native 32B

Model / precisionTaskOutputStepsTime
MiniMax H3 native 32BVideo with audio608×352, 39 frames20114.3 s
MiniMax H3 native 32BImage to video with audio608×352, 39 frames20175.5 s

Image and video setup on GitHub →

NVIDIA Image and Video Generation

Image and video models in ComfyUI on our NVIDIA workstation, with the denoiser on the RTX 3080 Ti and the text encoder and VAE on the RTX 3080, and no weights offloaded to system memory. LTX 2.3 is too large for one card, so its denoiser is split across both.Image and video models in ComfyUI, split across both cards.

Time from submitting the job to the saved image or video, in a fresh rerun of every model on November 5, 2026. Image rows have the model already loaded (the second image of a pair, with a new seed). Video rows are one clip each and include loading the model and encoding the prompt; starting the worker beforehand is not included. ComfyUI on CUDA, both cards in use; results vary with the model, settings and hardware. MiniMax H3 Image: November 15, 2026, a 3-bit GGUF build with its text encoder kept in system RAM; times are with the model loaded.

Samples under each model were generated on the same workstation during the November 4 and 5, 2026 benchmark runs.

Qwen-Image 2.1 Turbo INT8

Model / precisionTaskOutputStepsTime
Qwen-Image 2.1 Turbo INT8Text to image1024×102484.05 s
Qwen-Image 2.1 Turbo INT8Text to image2048×2048826.29 s

FLUX 3 Image Klein INT8

Model / precisionTaskOutputStepsTime
FLUX 3 Image Klein INT8Text to image1024×1024165.07 s
FLUX 3 Image Klein INT8Text to image2048×20481624.38 s

MiniMax H3 Image Q3_K

Model / precisionTaskOutputStepsTime
MiniMax H3 Image Q3_KText to image1024×102467.1 s
MiniMax H3 Image Q3_KText to image1536×1024612.0 s
MiniMax H3 Image Q3_KText to image4096×4096662.7 s

Wan 2.2 TI2V 5B FP16

Model / precisionTaskOutputStepsTime
Wan 2.2 TI2V 5B FP16Video512×288, 33 frames2051.56 sincludes loading and text encoding
Wan 2.2 TI2V 5B FP16Video1280×720, 33 frames2067.69 sincludes loading and text encoding
Wan 2.2 TI2V 5B FP16Video1920×1080, 33 frames20170.82 sincludes loading and text encoding

LTX 2.3 distilled Q4_0

Model / precisionTaskOutputStepsTime
LTX 2.3 distilled Q4_0Video with audio512×288, 33 frames884.93 sincludes loading and text encoding
LTX 2.3 distilled Q4_0Video with audio1280×720, 33 frames853.62 sincludes loading and text encoding
LTX 2.3 distilled Q4_0Video with audio1920×1080, 33 frames875.84 sincludes loading and text encoding

NVIDIA image and video setup on GitHub →

Ready to get started?

Tell us about your environment, we’ll design a deployment that keeps your data yours.

Get in Touch