Diffusers Model Loader

Diffusers Model Loader is the plain-Jane twin of ELLA Model Loader: same inputs, same output shape, minus the ELLA wrapping. It takes any standard SD 1.5 checkpoint via CheckpointLoaderSimple, converts it into a diffusers StableDiffusionPipeline, and hands you a DIFFUSERSMODEL that only Diffusers Sampler can consume. There's no T5, no 6 GB text encoder, no semantic-alignment magic - just your checkpoint, faithfully rebuilt in diffusers clothing.

Why would you want that? Because this pack is built around comparison. Diffusers Sampler runs this model with the stock CLIP text encoder, while ella_model_loader + ella_sampler run the same checkpoint with ELLA's T5-based encoder. Put both paths side by side with a fixed seed and the identical prompt, and you get a clean, apples-to-apples measure of what ELLA actually improves - and it does improve prompt adherence, though community testing in 2024 found it short of a properly fine-tuned SDXL model. This node is the baseline half of that experiment.

Mechanically it's the same conversion as the ELLA loader: it reads your checkpoint's state dict, converts the UNet and VAE into diffusers format, recreates the CLIP text encoder, and assembles a pipeline using the tokenizer configs shipped inside the pack. The one meaningful difference in the source: where the ELLA loader enables diffusers CPU offload, this path leaves everything on the GPU. That means the non-ELLA branch is the more VRAM-hungry one, which sounds backwards but is just the trade the author made.

Inputs and output are short:

• model (MODEL), clip (CLIP), vae (VAE) - the three standard outputs of a checkpoint loader.
• Output: diffusers_model (DIFFUSERSMODEL), wired into Diffusers Sampler.

Honestly, as a standalone node this is niche. If you just want to sample an SD 1.5 checkpoint in ComfyUI, the native KSampler is better integrated and faster. Its real value is the A/B pairing - and like every node in this pack, it's frozen at May 2024, superseded by Tencent's official ComfyUI-ELLA. Keep it for running old workflows, not for starting new ones.