Diffusers MV Sampler

This is the business end of MV-Adapter. Everything upstream - the pipeline loader, the VAE, the scheduler, the model makeup - exists to hand this node a fully-assembled PIPELINE. This is where you type your prompt, set your steps, and hit generate, and out comes the set of multi-view-consistent images: the same object, or scene, rendered from a ring of camera angles that actually agree with each other. That consistency is the whole point. Six views that look like the same thing from different sides are the raw material a 3D reconstruction or texturing step can turn into an actual model, which is why this pack got attention from the 3D-printing and game-asset crowd when it landed in late 2024.

How it works

MV-Adapter is an adapter bolted onto SDXL - it doesn't replace the base model, it teaches a frozen SDXL to generate several views at once that share lighting, identity, and geometry. So this node runs a fairly normal diffusion sample (prompt, negative prompt, steps, CFG, seed) but produces a batch of views in one shot instead of one image. Feed it a reference image and it works in image-to-multiview mode; leave that empty and it's pure text-to-multiview.

The inputs that matter

Most of the required inputs are the SDXL knobs you already know:

• prompt / negative_prompt - what to generate. Standard SDXL prompting applies; the pack ships a sensible default negative ("watermark, ugly, deformed, noisy, blurry, low contrast").
• num_views (default 6, max 12) - how many angles to produce. Six is the tuned default and pairs with the default azimuth_degrees of [0, 45, 90, 180, 270, 315] - front, front-right, right, back, left, front-left.
• steps (default 50) and cfg (default 7) - the usual quality/adherence dials. SDXL's community sweet spot is CFG 5–9, so the default sits right in range.
• seed - lock it while you tune, since multi-view results shift a lot run to run.

The optional inputs are where the pack's features live:

• reference_image (IMAGE) - the cleaned output of Image Preprocessor. Present = image-to-multiview; absent = text-to-multiview.
• controlnet_image and controlnet_conditioning_scale - for the scribble-to-multiview path, fed by Control Image Preprocessor once you've loaded a ControlNet into the pipeline.
• azimuth_degrees (LIST) - the camera angles. A View Selector node plugs in here to override the default ring when you only want, say, front and back.

The single output is IMAGE - the batch of views, which you send to a Save Image, or onward to whatever turns views into geometry.

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 and restart. First run auto-downloads SDXL plus the adapter weights from Hugging Face.

Where people get burned

VRAM. Because the base is SDXL, even with the memory-saving settings the README recommends (fp16 VAE, VAE slicing on) you're looking at roughly 13–14GB of GPU memory. This is not a 8GB-card node. If you OOM, that's the first thing to fix - not steps or resolution.

Second: this node is the last node in the chain, but it can't run on a bare pipeline. It needs the pipeline that's already been through Diffusers MV Model Makeup (which attaches the adapter, scheduler, and VAE). Wire a raw pipeline-loader output straight in and it won't behave.

Third, temper expectations. This is 2024-era tech, and the buzz around it faded through 2025 as dedicated image-to-3D models matured. The multi-view images are consistent and useful, but they're a front-end to a 3D pipeline, not a finished 3D model - you still need a reconstruction step downstream.