feat(anima): add prompt weighting support

[BMFreed]

Summary

Adds prompt weighting support to Anima text conditioning.

• Supports (text), (text:weight), nested groups, and escaped literal parentheses.
• Removes weighting markup before Qwen3 encoding.
• Maps weights onto T5-XXL tokens using tokenizer offset mappings.
• Passes the resulting weights through the existing Anima conditioning pipeline.

The implementation was written independently. ComfyUI was used only as a behavioral reference; no GPL-licensed source was copied or adapted.

Related Issues / Discussions

N/A

QA Instructions

• Verified that an unweighted prompt and the equivalent (text:1) prompt produce identical cleaned text, T5 token IDs, and weights.
• Verified that weighted text receives the requested T5 token weights.
• Verified that Qwen3 receives the cleaned prompt without weighting markup.
• Manually verified Anima generation in InvokeAI 6.13.8 using matching seed and generation settings.
• Ran the Anima test suite: 211 passed, 2 skipped.

Merge Plan

No special merge requirements.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated
• No redux slice changes
• No documentation changes required
• Updated Whats New copy (not a release PR)

[kappacommit]

Any chance you have before/after screenshots? My understanding is that Anima cannot actually support prompt weighting (in the traditional sense).

It does support limited prompt weighting of the form tag:weight eg: blue:1.2 but it is strictly a prompting/training concept and not actually a special function of the model.

I'd be curious to see a comparison of your feature on vs off, with the prompt weighting format described on the model page:

https://huggingface.co/circlestone-labs/Anima

Prompt weighting works, but needs a weight higher than typically used for SDXL. Example: "(chibi:2)"

[BMFreed]

Any chance you have before/after screenshots? My understanding is that Anima cannot actually support prompt weighting (in the traditional sense).

It does support limited prompt weighting of the form tag:weight eg: blue:1.2 but it is strictly a prompting/training concept and not actually a special function of the model.

I'd be curious to see a comparison of your feature on vs off, with the prompt weighting format described on the model page:

https://huggingface.co/circlestone-labs/Anima

Prompt weighting works, but needs a weight higher than typically used for SDXL. Example: "(chibi:2)"

Sure, here's a quick comparison. Long hair tag is amplified while the chibi tag is toned down. Without the fix the tag with weights often gets ignored entirely. You can check it out yourself.

[BMFreed]

My understanding is that Anima cannot actually support prompt weighting (in the traditional sense).

The Anima author explicitly confirmed that this is real inference-time prompt weighting, not merely a training/prompting convention. In ComfyUI the weights are applied by scaling the text-conditioning vectors produced by Qwen3 + the LLM adapter before they are fed into the diffusion model.

Author comment:
https://huggingface.co/circlestone-labs/Anima/discussions/135

And the implementation is visible here:
https://github.com/Comfy-Org/ComfyUI/blob/master/comfy/text_encoders/anima.py
https://github.com/Comfy-Org/ComfyUI/blob/master/comfy/ldm/anima/model.py