The verified 8 GB entry point
768² with float16, certified on the Turing sm75 RTX 2060 SUPER 8 GB.
- ≥ 8 GB VRAM
- 768²
- float16
- No Z-Image Turbo
COMPATIBILITY
Judge the GPU, model, and disk space first
The standard tier uses 1024² with bfloat16; the compact tier uses 768² with float16. Background removal and upscaling use separate runtimes.
What hardware is required?
Generative models require a supported 8 GB NVIDIA GPU for the compact tier. A 12 GB GPU enables the standard tier; Z-Image Turbo still requires 12 GB. BiRefNet and Real-ESRGAN are outside this tier system.
The complete 12 GB-and-up tier
1024² with bfloat16, including the current certified generative models and Z-Image Turbo.
- ≥ 12 GB VRAM
- 1024²
- bfloat16
- Includes Z-Image Turbo
COMPATIBILITY
Seven models and runtime boundaries
The app manages model identity, downloads, and runtimes; this table describes the certified scope only.
| Model | Capability | Hardware eligibility | Pinned recipe or runtime |
|---|---|---|---|
| Z-Image Turbo | Text to image | Standard tier only; 1:1, 4:3, 3:4 | Fixed 9-step recipe with 4-bit loading; not available on 8 GB GPUs. z-image-turbo |
| FLUX.2 Klein 4B distilled | Text to image / single-image edit | Standard and compact tiers | 4 steps with guidance 1.0, including reference-image editing. flux-klein-4b-distilled |
| BiRefNet general epoch_244 | Remove background | ONNX Runtime CPU | Returns transparent RGBA output plus a companion mask; not bound to the generative VRAM tiers. birefnet-general |
| Real-ESRGAN x4plus | Exact 4× upscale | ncnn Vulkan; tile 128 | Performs exact 4× upscaling and is not classified by the generative model VRAM tiers. real-esrgan-x4plus |
| Qwen-Image 2.1 | Text to image | Standard and compact tiers | Fixed 40-step recipe with 4-bit NF4 loading. qwen-image-2-1 |
| Qwen-Image 2.1 Turbo | Text to image | Standard and compact tiers | 8-step Turbo sampling for faster local generation. qwen-image-2-1-turbo |
| MOSS-Image-SFT | Text to image | Standard and compact tiers | Fixed 50-step recipe with 4-bit NF4 loading. moss-image-sft |
Z-Image Turbo
Text to image
Standard tier only; 1:1, 4:3, 3:4
Fixed 9-step recipe with 4-bit loading; not available on 8 GB GPUs.
FLUX.2 Klein 4B distilled
Text to image / single-image edit
Standard and compact tiers
4 steps with guidance 1.0, including reference-image editing.
BiRefNet general epoch_244
Remove background
ONNX Runtime CPU
Returns transparent RGBA output plus a companion mask; not bound to the generative VRAM tiers.
Real-ESRGAN x4plus
Exact 4× upscale
ncnn Vulkan; tile 128
Performs exact 4× upscaling and is not classified by the generative model VRAM tiers.
Qwen-Image 2.1
Text to image
Standard and compact tiers
Fixed 40-step recipe with 4-bit NF4 loading.
Qwen-Image 2.1 Turbo
Text to image
Standard and compact tiers
8-step Turbo sampling for faster local generation.
MOSS-Image-SFT
Text to image
Standard and compact tiers
Fixed 50-step recipe with 4-bit NF4 loading.
Model storage
Model storage
Installers do not include model weights. Each generative model is about 25–33 GB; reserve roughly 40 GB per model.
25–33 GB
Per generative model
~40 GB
Reserve per model
~150 GB
Full model set
- Plan about 150 GB for the full model set.
- Windows preparation requires at least 8 GiB available physical RAM and 6 GiB commit headroom.
- BiRefNet and Real-ESRGAN are not classified by generative VRAM tiers.
macOS
Apple Silicon arm64 with macOS 11 or newer.
Windows
Windows 10 22H2 or Windows 11 x64 with a supported NVIDIA GPU.
Hardware and storage facts come from the pinned source, model catalog, and current verification records without extrapolation.