Image
Run image generation, editing and upscaling in interactive or batch workflows.
World Model
Runtime
Hedra’s inference engine for image, video and world models. We tune kernels, numerical precision and scheduling for supported NVIDIA and AMD GPUs.
One stack for multimodal models.
Bring supported weights you own, license or build with Hedra.
Supported model workloads
Image
Video
Robotics + simulation
Shared engine
WMR
World Model Runtime
Vision kernels
Numerical precision
Scheduling
Execution
GPU execution
NVIDIA
AMD
Support agreed per workload
Video
Generate video from text or images, including for streaming workloads.
Robotics and simulation
Run vision-language-action policies and action-conditioned world models.
Inside the inference engine.
Built from our experience training and serving Hedra’s own models.
Kernels built for vision
We fuse operations for spatial-temporal attention, VAE encoding and decoding, and diffusion.
Precision tested against quality
We test how quantization affects image and video output quality.
Scheduling for diffusion
We batch requests and cache reusable computation around fixed-step generation.
NVIDIA and AMD support
We confirm model, GPU and precision support for your workload.
Faster on the same GPU.
Krea 3 GEMM kernel time on AMD Instinct MI355X.
PyTorch BF16 baseline
4.34 ms
Hedra custom MXFP6 kernel
1.54 ms
Profiler timings cover one GPU’s share of the MLP’s matrix multiplications. Both measurements use the same matrix shapes.
