TextGenerate could replace many custom nodes, but baulks at many LLMs (suspect feature request, but could be a bug)

### Feature Idea

The TextGenerate node is extremely helpful.

Even the most 'starred' LLM nodes have issues, such as a 50 second image>text using Qwen3-Vl-4b on a 5090, using nodes like https://github.com/1038lab/ComfyUI-QwenVL
...so it is great there is now core functionality. Thank you. But...

TextGenerate seems to be extremely limited - it would be useful if it could use more models.

- It can't use LLMs stored as shards (they need merging to a single safetensor)
- It can't use Qwen3-VL-4B, gives error: 'CLIPTextModel' object has no attribute 'generate'
- It can't use modified or abliterated versions of Gemma-3-12B or Gemma-4 (similarly node TextGenerateLTX2Prompt fails, but LTX's own Gemma 3 Prompt enhancer can use them)

On plus side,
- it can use Gemma 3 as a single .safetensor file
- it can use Qwen 3.5 as a single .safetensor file
- it can use every modified or abilterated version of Qwen3.5-4B / 9B i threw at it.

Is there potential to include Qwen3-VL-4B or not have it not baulk at albiterated Gemma3 or Gemma 4 models (the abliteration is not the fault here, as they do work with LTX's own Gemma 3 prompt enhancer node)

Fairly critical for Gemma particularly. Whilst it is amusing to be given 4 paragraphs of lecture about harm and animal cruelty in response, it is hardly necessary when requesting a prompt for a video in which the dog damages a lemon and a lady walks into the room and shouts at it. Never mind it's views on a battle scene.

Here is the error using Qwen3-VL-4B, and various others LLMs. (Oh, that's the dog ZIT made, about which Gemma got rather concerned).

And here is the error using any modified Gemma model (after which ComfyUI crashes with extensive log errors).

### Existing Solutions

None.

### Other

_No response_