Tiny Dream

Tiny Dream is a header-only Stable Diffusion inference library in C++, aimed at developers who need local text-to-image generation on machines with no NVIDIA GPU. Integration is two files — tinydream.hpp and stb_image_write.h — plus the pre-trained assets, and a single tinyDream class with 8 public methods. There is no OpenCV link; the only imaging dependency is the public-domain stb header used to write PNGs to disk. Version 2.197 runs on ncnn as its tensor backend, with a move to ggml on the published roadmap, and PixLab maintains a private x86-tuned fork of ncnn. Memory is the design constraint. ONNX quantization holds runtime at 1.7 GB to 4 GB for standard 512x512 output, and 5.5 GB is enough when Real-ESRGAN super-resolution upscales to 2048x2048 — that extra step is CPU-intensive and adds seconds. PixLab's own benchmarks for a 512x512 non-upscaled PNG, single seed, put an i9-9990XE at roughly 3.0–5.1s, an i7-13700 at 8.2–10.1s, and a Ryzen 9-5900X at 4.0–5.4s; reproduced timings vary on your machine. Control surfaces cover negative prompts for excluding gore or NSFW content, parenthesis parentheses weighting to boost words and square brackets to reduce them, guidance scale, sampling steps, and seed resizing to regenerate an image at a slightly different resolution. Output metadata can carry copyright notices or comments. Threading uses TBB with Intel MKL and AVX vectorization. The library writes nothing to stdout or stderr — you supply a log callback and route messages to a terminal, a file, or a socket. Support targets Stable Diffusion 1.x with the CLIP transformer for token embedding. Picking Tiny Dream means trading throughput and newer-model support for portability and a very small dependency graph; if you have a GPU and want speed or SDXL, Diffusers and ComfyUI are the obvious lanes. The PixLab pricing tiers below are the company's unified API plans, so confirm Tiny Dream licensing terms separately before you build around it.