DiffuEraser_Loader

This node is the assembly line for the pack's diffusion model. It has just two dropdowns, but behind them it's stitching together the whole video-inpainting pipeline: a VAE from your standard ComfyUI vae folder, a PCM LoRA from loras, a BrushNet mask-guidance branch and a temporal motion UNet from a pack-specific model folder, and the SD 1.5 config files the pack ships in its own repo. Everything in the DiffuEraser_Loader-less graph starts here.

The two inputs:

• vae - a dropdown of everything in ComfyUI/models/vae. The README's recommendation is sd-vae-ft-mse.safetensors, the usual stable VAE that most people already have for other workflows.
• lora - a dropdown of everything in ComfyUI/models/loras. This is where the PCM consistency-model LoRA goes, e.g. pcm_sd15_smallcfg_2step_converted.safetensors from wangfuyun's PCM_Weights (the sd15 folder). The node reads the filename to decide the step behavior - there are 2/4/8/16-step variants plus "Normal CFG" ones with real classifier-free guidance - so the LoRA you pick and the steps you set in the sampler should match.

Both dropdowns offer a literal "none" option, but you only want to see it if you're testing startup behavior. Without a VAE and a PCM LoRA the sampler will run with the wrong step budget and produce mush, or just fail.

How it works: the pack bundles the entire SD 1.5 config skeleton (sd15_repo - tokenizer, scheduler, model_index) inside the custom node folder, so the Loader builds a StableDiffusionDiffuEraserPipeline from those configs, drops your chosen VAE weights into it, then loads brushnet and unet_main subfolders from ComfyUI/models/DiffuEraser and finally applies the LoRA. Then it frees the GPU memory it can and hands you a ready model object.

Output: a single model output of the custom DiffuEraser_Loader type. It wires into exactly one place - the model input of DiffuEraser_Sampler.

What you have to download before this node does anything useful:

# ComfyUI/models/vae
sd-vae-ft-mse.safetensors            # from stabilityai/sd-vae-ft-mse
# ComfyUI/models/loras
pcm_sd15_smallcfg_2step_converted.safetensors   # from wangfuyun/PCM_Weights (sd15)
# ComfyUI/models/DiffuEraser/
#   brushnet/   (config.json + diffusion_pytorch_model.safetensors)
#   unet_main/  (config.json + diffusion_pytorch_model.safetensors)
#   ... both from lixiaowen/diffuEraser

Worth knowing: the README labels the pack's model folder "DiffuEraservae" - that's a typo. The code creates and reads ComfyUI/models/DiffuEraser, so put the BrushNet and motion-UNet files there. And the same folder also holds the ProPainter .pth weights used by the other half of the graph, which is convenient: everything heavy lives in one place. It's a chunky first download (multi-gigabyte UNet weights), so don't be surprised when the first run spends a while pulling from HuggingFace - use a download tool or the HF CLI rather than the browser if you can.