🏰 Load LDM Checkpoint

Loads a CompVis Latent Diffusion (LDM) checkpoint from ComfyUI/models/ldm. This is the original 2021 LDM architecture (Rombach et al.) with a BERT-based text encoder. The model operates in a 4-channel latent space with 8× spatial downsampling (f8 VAE). Outputs: MODEL / BERT / VAE. precision controls compute dtype: • fp32 — full precision (default, most stable); • fp16 — half precision (2× less VRAM, faster on modern GPUs); • bf16 — bfloat16 (better dynamic range than fp16, requires Ampere+). memory_mode controls VRAM usage: • keep — all components stay on GPU (fastest, highest VRAM); • auto — keep on GPU when VRAM permits, otherwise offload after stage; • offload — each component lives on GPU only during its stage; • cpu — entire inference runs on CPU (no VRAM required).