Community Forum

8GB is fine if you run with --lowvram or --medvram flags in ComfyUI and use FP8 quantized checkpoints. but once you layer ControlNet + IP-Adapter + 2 ...

16GB will fit 14B Q4 easily with around 45 tokens/sec on llama.cpp. but if u wanna use 16k context or 32B quant then definitely get the used 3090 24gb...