LDM Pipeline Loader

This is the node that lets MV-Adapter run on a checkpoint you already downloaded - a local SDXL .safetensors in your checkpoints folder - instead of pulling a diffusers repo off Hugging Face. For most people building serious multi-view workflows, this is the loader you actually want, because it's how you point the whole thing at Juggernaut, RealVis, or whatever community SDXL checkpoint gives you the look you're after, rather than being stuck on bare stable-diffusion-xl-base-1.0.

MV-Adapter deliberately ships two pipeline loaders. The diffusers one downloads by repo id; this LDM one loads local files the normal ComfyUI way. Same destination, different source.

How it works

"LDM format" just means the standard single-file checkpoint format the rest of ComfyUI uses, as opposed to the multi-folder diffusers layout Hugging Face serves. This node reads that file, sets up an SDXL pipeline around it, and hands you three things at once so you don't have to load them separately: the pipeline, the VAE baked into the checkpoint, and a default scheduler.

The inputs and outputs

• ckpt_name (enum) - the checkpoint file, from your checkpoints folder. This is the point of the node: pick your own SDXL model.
• pipeline_name (enum, default MVAdapterT2MVSDXLPipeline) - which pipeline class to build. Match it to your task: MVAdapterT2MVSDXLPipeline for text-to-multiview, MVAdapterI2MVSDXLPipeline for image-to-multiview, the SD variants for SD-2.1 checkpoints, or plain StableDiffusionXLPipeline for a non-adapter baseline. This has to line up with the adapter_name you choose in Model Makeup - t2mv here, t2mv adapter there.

Three outputs, all of which flow into Diffusers MV Model Makeup:

• PIPELINE - the model.
• AUTOENCODER - the checkpoint's VAE.
• SCHEDULER - a default scheduler.

That triple-output design is a nice convenience: one node feeds all three of Model Makeup's plumbing inputs. You can still override the VAE (with an LDM Vae Loader) or the scheduler (with the Scheduler Loader) if you want finer control - just don't wire this node's VAE output when you're supplying your own.

Installing the pack

ComfyUI Manager, search ComfyUI-MVAdapter, install, restart. Or: cd ComfyUI/custom_nodes && git clone https://github.com/huanngzh/ComfyUI-MVAdapter, then pip install -r requirements.txt, then restart. The _ldm-suffixed example workflows show this loader in place.

Where people get burned

Use an SDXL checkpoint (or SD 2.1 with the SD adapters). MV-Adapter is trained against SDXL and SD 2.1 only - drop in a Flux or SD 1.5 checkpoint and nothing lines up. Within SDXL, the well-supported niche is the mature realism and anime fine-tunes; that ecosystem is exactly where SDXL still lives in 2026.

Second, keep pipeline_name and adapter_name in sync across this node and Model Makeup. A text-to-multiview pipeline with an image-to-multiview adapter (or the reverse) produces nonsense.

Third, VRAM again - SDXL base means ~13–14GB even with the memory savings on. And note the checkpoint's built-in VAE may not be the fp16-fix version; if you hit black images or NaNs during decode, that's the classic SDXL-fp16 VAE problem, and the fix is to load sdxl-vae-fp16-fix through the LDM Vae Loader (with upcast_fp32 set to False) instead of using the baked-in one.