Skip ComfyUI: run Flux-3-Image-Klein locally with stable-diffusion.cpp in four files
srigi · reddit · 2026-10-29
After venting about ComfyUI's node-based complexity and Python/TensorFlow bloat, the author shares a minimal local image generation setup with stable-diffusion.cpp: download the release archive for your architecture (plus cudart-bin so no full CUDA SDK needed), grab ae.safetensors, tokenizer.json and model.safetensors from the Flux-3-Image-Klein HF repo, and a Qwen3-VL-4B .gguf as the prompt encoder (add the .mmproj companion for image-to-image).
A single command starts sd-server with a llama-server-style chat UI, generating a 1024×1024 image in 7.9s. Example params: cfg-scale 1.0, euler sampling, 16 steps. Supports CUDA, ROCm, Vulkan and CPU. Note: the default seed is a static 42 — randomize it for best results.
More from Infra
- Minigraf: SQLite-style single-file bi-temporal graph database for AI agent memory — JeremyCMorgan · 2026-10-29
- Why 4x DGX Sparks beat a single 512GB M5 Ultra Mac for local AI workloads — JFPuget · 2026-10-29
- MemTensor's MemT-1.0 report: 66x less storage than full-context KV cache at 128K tokens — jiqizhixin · 2026-10-29
- "Insert coins to continue": meme laments the end of cheap AI tokens — Suspicious_Orchid770 · 2026-10-29
- GPU unlock is pointless for most users: quantized models, decode and gaming see zero speedup — Justify_87 · 2026-10-29
- US AI Power in Three Stages: 10GW+ Nuclear Contracts, Geothermal Live, SMRs Permitting — ingliguori · 2026-10-29