HartsyInference: a pure C# generation backend for SwarmUI

Generate images, videos and music locally without managing Python, virtual environments, workflow servers or a pile of separate add-ons.

HartsyInference is a fast, all-in-one generation backend that runs inside SwarmUI. You keep the familiar model picker, prompt box, history, previews and generation controls. HartsyInference handles loading the model and creating the result in the same application.

It is in beta. Many model families and advanced tools work today, but exact feature support varies by model.

These results use the same request through the SwarmUI API on an RTX 4090. Each number is the warm median of three runs.

Hardware, checkpoint format, available memory, drivers and generation settings all change performance. Full test settings and newer results live in the HartsyInference performance guide.

Install it from Server > Extensions, or clone the repository into SwarmUI/src/Extensions/ by hand. Extensions are compiled, so restarting by itself is not enough: run update-windows.bat or update-linuxmac.sh from the main SwarmUI folder, then start SwarmUI.

HartsyInference ships as a package dependency of this extension. You do not install the engine separately.

To add the backend, open Server > Backends, turn on Show Advanced, select HartsyInference (Pure C# Inference), keep ComputeBackend, GPU_ID and LowVram at their defaults, and save. For a normal one-GPU setup, leave every field ending in GpuId empty.

HartsyInference uses SwarmUI's normal model folders and model picker. Your model root is shown under Server > Server Configuration > Paths > ModelRoot. Place a checkpoint in the matching folder and refresh the model list.

For a straightforward first image, use a supported turbo or schnell checkpoint that fits your GPU. Krea 2 Turbo and Z-Image Turbo are fast choices on higher-memory NVIDIA cards. Some model families need companion files such as a text encoder or VAE; HartsyInference downloads known companions when it can, and SwarmUI shows what is missing when it cannot.

The finished result appears in the normal SwarmUI history. Select it to view the full image, prompt, settings, seed and generation details, exactly as with any other backend. Nothing about the output path changes: the same metadata, the same sharing, the same history.

While it runs, you keep the controls you would expect. HartsyInference supports live previews, cancellation, variation seeds and early step stopping. If a model cannot use a feature you asked for, the backend says so and names the problem rather than quietly ignoring the setting, which is the difference between a slow run and a run that was never going to do what you wanted.

Images. Stable Diffusion 1.5, SDXL, SD3, SD3.5, FLUX.1, FLUX.2, Qwen-Image, Z-Image, Krea 2, Chroma, AuraFlow, Anima, F-Lite, Ideogram 4, HunyuanImage, ERNIE-Image and Boogu.

Editing and repair. Start from an existing image and control how much it changes, paint a mask to replace or repair an area, refine one model's result with another, make seamless images for repeating textures, and use supported reference-image models for instruction-based edits.

Guided composition. Pose, depth, edges, line art, soft edges, scribbles, normals and segmentation maps, plus IP-Adapter variants that carry style, facial features or identity from a reference. Supported ControlNets and LoRAs stack with individual strengths, and each control appears only when the selected model can use it.

Video. Wan 2.x, Wan VACE, Wan Animate, Wan speech-to-video, LTX-Video, LTX-2, HunyuanVideo, Kandinsky 5 and MiniMax H3, with controls for frame count, frame rate, format, image-to-video, reference media, trimming, boomerang playback, audio and SeedVR2 restoration.

Music. ACE-Step checkpoints generate music from the normal Generate tab, with music controls appearing once you choose a compatible model.

To run separate jobs at the same time, add one HartsyInference backend per GPU and give each a different GPU_ID. SwarmUI then sends different generations to different cards.

To split one model across two GPUs, a single backend can use a second card for part of the work: TextEncoderGpuId moves prompt processing, VaeGpuId moves image or video encoding and decoding, DitShardGpuId splits a supported image or video model to combine their memory, LmShardGpuId splits a supported language or audio model, and CfgParallelGpuId runs two guidance passes at once when both cards can hold the model.

Sharding is mainly a way to gain memory. It does not always improve speed.