Confused about Embeddings in Minimax H3

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Your question

I'm testing the embeddings from:
https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/embeddings
I'm going to focus mainly in minimaxh3_bullet_time.safetensors which is the one that is crealy working for me. When I use the embedding with a reference the embedding seems to work, but when I try pure text2img it seems totally ignored.

https://github.com/user-attachments/assets/4613ffaa-f3fd-40a7-a98e-3138b67c1495

Prompt, seed and everything the same in both cases, the only thing that changes is adding the reference image. I've seem the same behaviour in other embeddings like the kissing to the camera.

-------------
embedding:minimaxh3_bullet_time,

a woman walking over stone bridge revealing at the end a giant stone structure similar to a bird made of giant stone blocks.
Sound: Instrumental Vivaldi's classical music orchesta.
----------------

### Logs

```powershell

```

### Other

_No response_

Start by reproducing the two workflows with minimaxh3_bullet_time.safetensors from the MiniMax-H3 embeddings page, keeping the prompt and seed fixed while comparing text2img with reference-image generation. Determine whether the embedding is expected to affect pure text2img; done means explaining the difference or identifying a reproducible implementation problem. No source file or test is named, and the empty logs provide no narrower entry point.

Written by the indexing model from the issue text.