View Selector

By default MV-Adapter generates a full ring of six views. This node lets you pick a subset - say just front and back, or front/right/back - instead of always producing all six. It's a small quality-of-life node added in January 2025: six on/off toggles, one output, and it plugs into the sampler to say "these are the angles I actually want."

Why bother? Sometimes you don't need the whole turntable. If you only want a front-and-back pair for a flat asset, or three-quarter coverage for a simple object, generating six views wastes compute and - more importantly - the adapter's beta model is measurably better at small view sets than at the full six. This node is how you take advantage of that.

How it works

Each of the six toggles corresponds to one camera angle in the standard ring. Flip the ones you want on, leave the rest off, and the node emits a list describing that selection. That list drives which azimuths the Diffusers MV Sampler actually renders.

The inputs and output

Six boolean toggles, all defaulting to true (i.e. the full ring), named for their angles:

• front_view, front_right_view, right_view, back_view, left_view, front_left_view

Turn off the ones you don't want. Per the README's rough testing, the beta image-to-multiview model handles certain subsets especially well: 2 views (front & back), 3 views (front, right, back), and 4 views (front, right, back, left).

The output is a LIST - it wires into the azimuth_degrees input on the Diffusers MV Sampler, overriding the default six-angle ring with your chosen subset.

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 workflow to copy is i2mv_sdxl_ldm_view_selector.json, which wires the selector into an image-to-multiview graph.

Where people get burned

Two things straight from the README, and both are easy to miss.

First, the beta adapter. To get good results with view selection, the README says to set adapter_name in Diffusers MV Model Makeup to mvadapter_i2mv_sdxl_beta.safetensors - the beta image-to-multiview model is the one tuned for partial view sets. Use the non-beta adapter with a reduced selection and the results are noticeably worse. Match the model to the mode.

Second - and this one genuinely confuses people - when you use the View Selector, the num_views attribute is ignored. It's in the README in plain terms: with view selection active, num_views does nothing and can be left alone. So if you've wired in a selector and you're fiddling with num_views trying to change the count, that's why nothing happens. The toggles here are what control the view set, not the number field.

Third, keep your selection sane. The beta model was validated on the front/back, front-right-back, and front-right-back-left subsets; oddball combinations (say, left and front-left only) aren't what it was tuned for and may come out weaker. Stick close to the tested groupings unless you're just experimenting.