Guide
How much VRAM does MOSS-Image need?
MOSS-Image has two large parts: a 5B diffusion model (10.2 GB in bf16, 5.4 GB in int8) and a Qwen3.5 9B text encoder (19.3 GB in bf16, 9.8 GB in int8, 7.3 GB in w4a8). Loading both in bf16 takes about 30 GB of weights alone. With the int8 model, a compressed text encoder and offloading, VRAM requirements can be reduced, but compatibility still depends on your GPU, system RAM and workflow. Official support is NVIDIA CUDA only.
File sizes at a glance
| Part | bf16 | int8 | w4a8 |
|---|---|---|---|
| Diffusion model (SFT or Base) | 10.2 GB | 5.4 GB | — |
| Text encoder (Qwen3.5 9B) | 19.3 GB | 9.8 GB | 7.3 GB |
| VAE | 0.34 GB (small: 0.25 GB) | — | — |
These are file sizes from the Comfy-Org repository, not measured peak VRAM. Actual use also depends on resolution and on whether parts are moved off the GPU between steps.
Which setup fits your GPU (estimate)
- 24 GB or more: bf16 diffusion model, with the text encoder in int8 or offloaded after encoding.
- 12–16 GB: int8 diffusion model plus the w4a8 text encoder, with offloading enabled.
- Under 12 GB: use the official “Sequential” memory mode or community GGUF builds, and expect slow generation.
These configurations are our estimates from file sizes, not a benchmark. We will update it with measured numbers.
The three official memory modes
In the official Gradio app, Advanced settings offers three modes:
- GPU resident: the whole model stays on the GPU. Fastest, needs the most memory.
- Component: parts move between CPU and GPU as needed.
- Sequential (fallback): lowest memory use, slower.
If you run out of memory, choose Sequential or reduce the image size.
Low-VRAM and quantized options (community)
- GGUF: realrebelai/MOSS-Image_GGUFs (GGUF and W4A8) and leejet/MOSS-Image-SFT-GGUF (for stable-diffusion.cpp).
- Fewer steps: lodestones/MOSS-Image-SFT-8-steps-dmd (8 steps, no CFG) and akhaliq/MOSS-Image-SFT-4-steps (4 steps, CFG 1, a LoRA for the stock Euler sampler).
These are community builds, not official releases. Quality can differ from the full model.
Can I run MOSS-Image on a Mac?
Not officially; the official pipeline requires CUDA. A community member reported working SFT inference on an M1 Max with 32 GB of unified memory, using an Apple Silicon (MPS) adaptation proposed on GitHub. That does not establish compatibility with 16 GB Macs; we have not tested Macs ourselves.
No suitable GPU?
Try MOSS-Image AI in your browser. Generation runs in the cloud, so you do not need a local GPU. Eligible visitors can try one free Standard image; see the free-use guide for limits.
Few-step builds target speed and do not guarantee lower peak VRAM. See the Comfy-Org files and official memory-mode documentation.
Questions
Why is the text encoder bigger than the image model?
MOSS-Image uses a 9B Qwen3.5 language model to understand long prompts of up to 4,096 tokens. The encoder interprets prompts; it does not guarantee accurate text in the output.
Does a bigger image need more VRAM?
Yes. The official advice is to pick a lower memory mode or reduce image size when memory runs out.
Is int8 much worse than bf16?
There is no official comparison yet. Test on your own prompts, especially ones with small text.
How much VRAM do I need to train a LoRA?
The community KohakuMossTrainer reports adapter training with as little as 6 GB of VRAM using its memory-saving options.
Does it run on AMD GPUs?
The official pipeline lists NVIDIA CUDA only.
Turn your idea into an image
Describe your idea, generate an image, then refine the text or layout if needed.