BD TRELLIS.2 Get Conditioning

TRELLIS.2 normally makes you juggle two things to get from an image to a 3D shape: a model configuration (which resolution, which attention backend) and the DinoV3 visual features extracted from your image. BD TRELLIS.2 Get Conditioning fuses both into one node - the model config is embedded in the conditioning it outputs, so there's no separate config node to keep in sync. Feed it an image and a mask, and it hands you everything the downstream shape/texture nodes need, plus a preprocessed image so you can see what the model actually saw.

What it is

The entry point of the BrainDead TRELLIS2 flow. It loads the TRELLIS.2 model configuration (default microsoft/TRELLIS.2-4B), extracts DinoV3 features from your image, and packs the whole thing into a TRELLIS2_CONDITIONING output. Wire that one output to BD TRELLIS.2 Image to Shape and to BD TRELLIS.2 Shape to Textured Mesh, and the single-workflow version of the pipeline is done.

How it works

The required inputs are image and mask - the subject and its silhouette, which is what tells the model what to reconstruct and what to ignore. Then the meaningful combo:

• resolution - 512 (fast, ~15s, lower detail, good for shape), 1024_cascade (balanced ~30s, the default), 1536_cascade (highest detail ~45s, best for texture). This is the primary quality-vs-speed dial.
• attn_backend - flash_attn default, with xformers / sdpa fallbacks if you don't have flash attention compiled.
• vram_mode - keep_loaded (fast if you'll run again) or cpu_offload (for stacking with other models).
• include_1024 - extract the 1024px features the cascade modes need. Leave it on unless you're doing pure-512 work and want to save a little time.
• background_color - the color used when preprocessing composites the subject; gray default, and it matters for edge quality on light/dark subjects.

Outputs: conditioning and preprocessed_image - that second one is your free sanity check that the crop and background are right before you burn 30 seconds of sampling.

The dual-workflow trick

The node is built around being used twice. The recommended setup from the README:

1. First node at resolution=512 → feed its conditioning to Image to Shape for a fast shape (~15s).
2. Second node at 1536_cascade, optionally with a different image → feed its conditioning to Shape to Textured Mesh as texture_conditioning.

Result: fast shape generation with a high-detail texture voxelgrid. Want to go further - traced outline for shape, clean image for texture? That's the BD TRELLIS2 Dual Conditioning info node's territory (it's a docs node, not a processor). The point is the conditioning from this node isn't a one-shot; it's designed to be split across the two stages.

Gotchas

The real dependency: this node imports the trellis2 package from the separate ComfyUI-TRELLIS2 pack - the BrainDead nodes are convenience wrappers, not a standalone implementation. Install ComfyUI-TRELLIS2 first or the node reports TRELLIS2 unavailable. The model itself (TRELLIS.2-4B) downloads on first use and is heavy - budget VRAM accordingly; vram_mode=cpu_offload is your friend on a tight card, and max_tokens on the downstream shape node is the other VRAM lever.

Installing it

ComfyUI Manager: search "BrainDead" → install. Manual:

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

Also install ComfyUI-TRELLIS2 (via Manager or git clone into custom_nodes/). Restart, find this under 🧠BrainDead/TRELLIS2. Start at 1024_cascade with a clean subject mask, verify the preprocessed_image looks right, and you're one wire away from a shape.