BD TRELLIS.2 Multi-Image Conditioning

The single biggest reason a one-image-to-3D model gets the back of an object wrong is that it only ever saw one side of it. This node is the fix: it lets you hand TRELLIS.2 several photos of the same thing - front, side, back, whatever you've got - so the shape is built from actual views instead of guesses. It's the difference between "this chair has a mystery rear end" and "this chair was a chair from every angle."

It's part of the 🧠BrainDead/TRELLIS2 family in the ComfyUI-BrainDead pack, and it's the multi-view sibling of BD TRELLIS.2 Get Conditioning. Where that node encodes a single image, this one takes a whole batch and concatenates their features so the model attends to all of them at once - the same trick TRELLIS 1's multi-image mode used.

How it works

Internally it runs DinoV3 over every image in your batch, extracts the feature tokens, and concatenates them into one TRELLIS2_CONDITIONING object. When you feed that into BD Image to Shape or BD Shape to Textured Mesh, the model sees all views simultaneously rather than conditioning on a single frame. You get more accurate geometry - and, just as usefully, fewer hallucinated details on the unseen sides.

The two inputs that matter are both batched:

• images - your reference views, batched. If you only have one image, you don't need this node; use the plain Get Conditioning. The tooltip says it plainly: batched images with B > 1.
• masks - a corresponding foreground mask per image, so TRELLIS.2 knows what's object and what's background.

Then there are the usual TRELLIS.2 knobs:

• resolution - 512 (fast, low detail, fine for shape), 1024_cascade (the balanced default), 1536_cascade (highest detail, best when the output is feeding texture generation).
• attn_backend - flash_attn by default; if you get import errors, sdpa is the reliable fallback and xformers is the middle ground.
• vram_mode - keep_loaded (fast, holds the model in VRAM) or cpu_offload (slower, but you can stack it next to other models).
• include_1024 - extract the 1024px features the cascade modes need. Leave it on unless you're only ever running 512.
• background_color - what the preprocessed crops are composited onto; gray is the default and usually fine.

Outputs: conditioning (wire it into Image to Shape or Shape to Textured Mesh) and preprocessed_images, a handy visual of what the node actually cropped and cleaned before encoding - check it once if the results look off, because a bad mask shows up there immediately.

Installation

The node ships in ComfyUI-BrainDead. Easiest route: ComfyUI Manager → search "BrainDead" → install. Manual is the same as any custom node:

cd ComfyUI/custom_nodes
git clone https://github.com/BizaNator/ComfyUI-BrainDead
cd ComfyUI-BrainDead
pip install -r requirements.txt

Here's the gotcha nobody warns you about: the BrainDead TRELLIS2 nodes are not self-contained. They import the trellis2 package and, if it's missing, fail with an ImportError telling you to install ComfyUI-TRELLIS2. That other pack is what actually holds the model weights and downloads the multi-GB TRELLIS.2-4B model on first use. Install it too, or every TRELLIS2 node in this pack just errors out at runtime.

Common issues

• "trellis2 not available" - install ComfyUI-TRELLIS2 as above. This one's the whole troubleshooting section, honestly.
• Nothing changes when you add views - check you're actually feeding a batch with B > 1. A single image is a degenerate multi-view input; use a Batch node to stack your shots.
• OOM or slow runs - switch attn_backend to sdpa if flash attention isn't compiled for your torch build, and drop vram_mode to cpu_offload on cards under ~12GB. The 3D-gen corner of ComfyUI is where the install fight lives; this pack keeps it light, but TRELLIS.2 itself is still a 4B model and wants room.

One honest note, and it applies to all of this pack's 3D output: TRELLIS.2 gives you lovely surface quality and the MIT license is refreshing, but the topology underneath is machine-generated triangle soup. For a static prop or a render, brilliant. For something that has to rig and deform in a game, budget for a retopo pass.