Quantizing an image model does not remove the cost of its supporting components: the diffusion weights may fit in a few gigabytes, while the accompanying text encoder remains substantial. This repository’s main value is giving users several precision and runtime choices for the same 7B, 8-step Turbo model, rather than presenting one compressed checkpoint as a universal low-memory solution.
Key Capabilities
- The available GGUF options range from Q4_K_M at about 4.05 GB to Q8_0 at 7.59 GB, letting users trade weight size against precision.
- Safetensors variants include FP8, INT8 ConvRot, and NVFP4, while MLX files support 4-, 6-, and 8-bit workflows. This makes the checkpoint adaptable to ComfyUI, command-line inference, and MLX-based setups.
- The underlying Turbo model uses eight denoising steps and guidance scale 1 by default. Its short sampling schedule is the reason it can generate quickly; using a generic sampler schedule may not reproduce the intended behavior.
- Companion files matter for memory planning: the supplied INT8 text encoder is about 9.35 GB and the BF16 VAE about 0.68 GB. Model file size alone is not a reliable estimate of total runtime memory.
Who It Suits
A good fit for people who already run local image-generation workflows and want to compare quantization formats or use the same model across ComfyUI, GGUF-capable runtimes, and MLX. Look elsewhere if you need a small all-in-one download, a managed inference service, or a model with an integrated safety filter; the quantized diffusion weights are only one part of the setup, and the base model does not include a built-in safety checker.
