Skip to content

The v4 engine

The v4 engine is Easy Diffusion's experimental engine, based on sdkit3.

How to use

  1. Open the Settings tab in Easy Diffusion's UI.
  2. Select v4 in the Engine to use (Backend) option. See Engines for the other engines.
  3. Press Save and restart Easy Diffusion.

You can also choose the Backend platform below it: Auto (default), or Vulkan (experimental).

Fails to start?

If v4 fails to start, please copy all the logs from the command prompt (or terminal) window, and attach it in a post on our server.

Important: Please don't leave out any lines from the logs, since they contain useful information required for investigating the failure.

Details

It uses stable-diffusion.cpp and ggml under-the-hood, and produces optimized, lightweight builds for the target hardware.

The main benefits of the v4 engine are:

  1. Very lightweight: less than 100 MB install footprint, compared to 3 GB+ for Forge and other PyTorch-based engines.
  2. Much better for AMD/Intel/Integrated users: avoids the hot mess of ROCm and DirectML, by using a reliable Vulkan backend (that's also used in llama.cpp).
  3. Opportunity for even faster image generation in the future: this currently uses stock sd.cpp, which has room for further optimization.
  4. Support for older GPUs: Vulkan supports older GPUs, especially older AMD GPUs unsupported by ROCm/PyTorch.

Supported features

  • Reference Images: an input section in the UI where you can attach one or more images, for vision-guided models (e.g. image editing models like Qwen). The model looks at these images in addition to your prompt. This input is only available on the v4 engine.
  • Scheduler: the Scheduler dropdown from Image Settings is available; see Samplers for the options.
  • Distilled Guidance: available for models that use it (e.g. Flux); see Flux and other new models.
  • Text Encoder: available, for models that need a separate text encoder file (including vision mmproj files).
  • Tokenizer: available, for models that need an external tokenizer.json. See Text encoders and tokenizers.

Limitations and caveats

  • The v4 engine is very experimental. It may fail to start or produce errors on some hardware, please report with full logs (see above).
  • Some models, samplers and upscalers supported by the v3.5 engine are not available on v4 yet.
  • Face correction and upscaling options are more limited than on the other engines.
  • If v4 doesn't work well for you, switch back to v3.5 in the Settings tab (Save + restart).