Diffusers MV Pipeline Loader

This is usually the first node in an MV-Adapter graph. It loads the base SDXL model by Hugging Face repo id - auto-downloading it if you don't have it cached - and sets up the diffusers pipeline that everything else builds on. If you want the zero-setup path where you don't have to hunt down and place checkpoint files yourself, this is it: type a repo name, and the pack fetches the weights on first run.

It's the diffusers-format twin of the LDM Pipeline Loader. This one pulls from Hugging Face; the LDM one loads local files. Pick based on whether you'd rather auto-download or use checkpoints you already have.

How it works

MV-Adapter is built on diffusers, and diffusers models load by repository id rather than by filename. This node takes a repo id, resolves it to a full SDXL pipeline (downloading and caching the weights the first time), and outputs the pipeline plus its VAE and a default scheduler - all three of which feed into Diffusers MV Model Makeup, where the adapter actually gets attached.

The inputs and outputs

• ckpt_name (STRING, default stabilityai/stable-diffusion-xl-base-1.0) - the Hugging Face repo id of the base model. The default is stock SDXL, which is the safe starting point. You can put any diffusers-format SDXL repo here, but note the string is a repo id, not a local file - that's the whole difference from the LDM loader.
• pipeline_name (enum, default MVAdapterT2MVSDXLPipeline) - which pipeline class to build: MVAdapterT2MVSDXLPipeline for text-to-multiview, MVAdapterI2MVSDXLPipeline for image-to-multiview, the SD variants for SD-2.1 bases, or plain StableDiffusionXLPipeline for a baseline. Keep this aligned with the adapter_name in Model Makeup.

The three outputs - PIPELINE, AUTOENCODER, and SCHEDULER - all wire into Model Makeup. One node, three plumbing connections handled.

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 diffusers-path example workflows are t2mv_sdxl_diffusers.json and i2mv_sdxl_diffusers.json.

Where people get burned

First run downloads. Because the weights come from Hugging Face on demand, the first time you queue a graph the node stalls to fetch several gigabytes of SDXL. That's not a hang - it's the download. You need a connection and some patience; after that it's cached.

Second, the repo has to be a diffusers-format SDXL repo. A single-file .safetensors checkpoint you downloaded from CivitAI won't load here - that's what the LDM Pipeline Loader is for. If your favorite community checkpoint only ships as a single file, use the LDM path instead.

Third, keep pipeline_name matched to your adapter choice downstream, and remember the base is SDXL - so the usual ~13–14GB VRAM floor applies, and stock stable-diffusion-xl-base-1.0 is deliberately plain. For nicer results, load a stronger diffusers SDXL fine-tune here, or switch to the LDM path and point it at a community checkpoint. The multi-view consistency comes from the adapter; the look still comes from whatever base you feed it.