Show and tell: Vison, a Windows desktop app built on stable-diffusion.cpp (Vulkan, FLUX / Z-Image / Wan)

[JayRGadekar]

Thanks for stable-diffusion.cpp. I've built a desktop app on it, Vison: https://github.com/JayRGadekar/Vison (MIT). It wraps a C++ server around stable-diffusion.cpp and vision.cpp, with an Electron UI.

What it does with your work: image gen (SDXL Turbo, Z-Image Turbo, FLUX.1 Schnell, Qwen-Image), video (Wan 2.1/2.2, HunyuanVideo 1.5, LTX-2.5 via the upstream patch, MiniMax H3), image-to-image, upscaling, and "paste a Hugging Face link and it picks the right encoder/VAE for the family".

A few things you might find useful, all documented in PATCHES.md:

• city96's FLUX.1-schnell-gguf names its tensors bare (double_blocks.0...), so get_sd_version() rejects it as an unknown architecture; I patched detection locally. Would that be welcome as a PR?
• The streaming-VRAM budget bug (the clamp being undone by the high-water-mark ratchet, then an allocator abort). I see it was fixed upstream in 1706b32, and I'll drop my patch when I move the pin.
• A packaging constraint: vision.cpp bundles its own ggml fork, so I can't bump stable-diffusion.cpp past 487de75 without a symbol clash on the shared ggml target.

Status: one dev, one 6GB NVIDIA laptop GPU, Windows only. Qwen-Image, Wan 2.2 A14B, HunyuanVideo 1.5, LTX-2.5 and MiniMax H3 are wired up but untested for lack of VRAM. Happy to share logs from the 6GB card if useful.

[wbruna]

city96's FLUX.1-schnell-gguf names its tensors bare (double_blocks.0...), so get_sd_version() rejects it as an unknown architecture;

It's this patch?

https://github.com/JayRGadekar/Vison/blob/main/third_party/patches/0001-sd-cpp-vison.patch

It should work without it, as long as you pass it as --diffusion-model (which adds the prefix model.diffusion_model.). The working gguf I have here has that same structure.

A packaging constraint: vision.cpp bundles its own ggml fork, so I can't bump stable-diffusion.cpp past 487de75 without a symbol clash on the shared ggml target.

Releases past master-883-137f740 support building without stable-diffusion.cpp changes; search for SD_USE_UPSTREAM_GGML.