ELLA Combine Embeds

If you've used T5 Text Encode #ELLA or Combine CLIP & ELLA Embeds, you've noticed they output something called ELLA_EMBEDS rather than CONDITIONING. That's a dict - a bag of named embedding tensors (T5 states, CLIP states, pooled outputs) that gets turned into real conditioning later. ELLA Combine Embeds is the glue for that bag: it merges two of them into one.

Mechanically it's a dictionary merge - {**embeds, **embeds_add} - with a warning if both bags share a key (because there are some same keys, one of them will be overwritten). That's the whole node. It doesn't average embeddings, it doesn't concat them, it just unions the contents so a downstream ELLA Encode sees everything.

When you'd actually use it

• You encoded two different prompts with T5 (say, one for the subject, one for style) and want them in the same bag before encoding.
• You're composing from multiple sources - a T5 branch and a CLIP branch - and need to carry them together.
• You want to layer an ELLA_EMBEDS bag produced elsewhere (chained through T5 Text Encode's optional embeds input) with additional content.

Honestly, most ELLA workflows never need it - the pack's composition story is mostly T5 Text Encode → (optionally Combine CLIP & ELLA Embeds) → ELLA Encode. This node is for the cases where you're building conditioning programmatically or mixing more than two sources, and it's the piece you reach for instead of trying to jam extra data into an existing bag by hand.

Inputs and output

• embeds - the base ELLA_EMBEDS bag.
• embeds_add - the bag to merge in. Its entries win on key collisions, which is why the warning exists.
• Output - a merged ELLA_EMBEDS bag.

Same as every node in this pack: install via ComfyUI Manager or git clone https://github.com/TencentQQGYLab/ComfyUI-ELLA into custom_nodes (pip install -r requirements.txt), with the ELLA model in models/ella and FLAN-T5 XL in models/ella_encoder. There's no model logic here at all, so nothing else to configure.