BD Cache Trellis2 Conditioning

A TRELLIS2 workflow isn't one slow step - it's several, and the first one is the conditioning pass that runs your image through the model's preprocessing to build the prompt/conditioning bundle. It's not the flashy stage, but it's a real chunk of time, and it runs every single time unless you stop it. BD Cache Trellis2 Conditioning is the stop.

The node description tells you exactly where it goes: "Place AFTER Trellis2GetConditioning node." You wire the TRELLIS2_CONDITIONING output in, and on the first run it's cached; on every run after that, the lazy-evaluation machinery returns "no inputs needed," so ComfyUI skips the GetConditioning node (and everything upstream of it - the image loading, any preprocessing) entirely. The conditioning bundle loads from disk and flows into the shape stage exactly as if it had just been computed.

This is the first of three TRELLIS2-specific caches in the pack, and they're designed to stack: condition → shape → texture, each one cachable so you can work on any segment of the pipeline without paying for the earlier ones. If you're iterating on the shape or texture parameters, you want this node in the graph even if you're not planning to touch the image.

The inputs

• conditioning - the TRELLIS2_CONDITIONING output from Trellis2GetConditioning (or your TRELLIS2 pack's equivalent).
• cache_name - label (default trellis2_cond).
• seed - invalidation key, hashed into the filename. Change it when the source image or conditioning settings change.
• force_refresh - True to overwrite and re-run the preprocessing.
• name_prefix - optional path prefix under output/BrainDead_Cache/.

Outputs: conditioning (same type back out) and status (STRING).

The workflow shape

[Trellis2GetConditioning] → [BD Cache Trellis2 Conditioning] → [Trellis2ImageToShape]
                             cache_name: "source_01_cond"
                             seed: 42

First run preprocesses and caches. Every run after that skips the whole preprocessing branch. It's the cheapest of the three TRELLIS2 caches to hit, and it makes the shape cache downstream work better too - the conditioning is now stable on disk, so the shape stage has a deterministic input.

Installing it

Part of BizaNator/ComfyUI-BrainDead (BrainDeadGuild). ComfyUI Manager → search "BrainDead" → install, or:

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

Restart; it's under 🧠BrainDead/TRELLIS2. You still need a TRELLIS2 node pack installed separately - this node caches the conditioning type, it doesn't compute it.

Gotchas

The usual seed rule, plus one TRELLIS2-specific warning: this node only knows the seed hash, not the image. Swap the source image, forget to bump the cache seed, and you'll keep feeding the stale conditioning into the shape stage - which produces a very confusing "why is it still generating the same shape" moment. When you change the input image, change the cache seed or flip force_refresh for one run. And since conditioning depends on the exact TRELLIS2 model version you're using, update the model → invalidate the cache.