ComfyUI MOSS-Image Troubleshooting & Error Fix Guide
Instant solutions for common MOSS-Image startup crashes. Find tested fixes for CUDA Out of Memory (OOM) on 8GB/12GB graphics cards, missing native layout nodes, and Qwen3.5-9B text encoder tokenizer errors.
Red node boxes in ComfyUI canvas showing 'Node not recognized' or 'Missing Custom Node' upon dragging workflow PNG.
Your ComfyUI installation is running an older commit before native MOSS-Image PR integration, or custom nodes were placed in the wrong folder.
- Navigate to your ComfyUI root directory in terminal.
- Run `git pull` to fetch the latest upstream nodes and PR patches.
- If you are using manual custom nodes, clone `ComfyUI-MOSS-Image` into `ComfyUI/custom_nodes/`.
- Restart ComfyUI backend server.
cd ComfyUI && git pull && pip install -r requirements.txtTerminal crashes with `torch.cuda.OutOfMemoryError: Tried to allocate 4.20 GiB (GPU 0; 8.00 GiB total capacity)` during prompt encoding.
Qwen3.5-9B is a 9-billion parameter model. In full BF16 precision, loading it alongside the 10.2GB MOSS diffusion weights exceeds 18GB VRAM.
- Launch ComfyUI with `--lowvram` flag: this forces automatic CPU offloading for text encoding.
- Replace the 18GB BF16 text encoder with the FP8 quantized version (`qwen_3.5_9b_fp8.safetensors`, only 4.8GB).
- Enable `--cpu-vae` to decode the final image in system RAM rather than GPU VRAM.
- If running on an 8GB card and encountering extreme slowdowns, consider cloud GPU instances.
python main.py --lowvram --fp8_e4m3fn --cpu-vaeUniversal 3-Step Clean ComfyUI Launch Command
If you are encountering frequent memory crashes or slow generation on cards with ≤ 12GB VRAM, use these startup parameters:
python main.py --lowvram --fp8_e4m3fn --cpu-vae --preview-method auto💡 Tired of Hardware Bottlenecks & CUDA Crashes?
Don't spend hours tweaking Python packages and pagefile sizes. Rent an RTX 4090 instance on RunPod or MimicPC for $0.29/hr and generate without limits.