Expanded IP Adapter modes.

[hipsterusername]

This pull request introduces enhancements to the IP-Adapter functionality, including the addition of two new methods (style_strong and style_precise) and corresponding updates across the backend and frontend to support these methods. Key changes include updates to data models, invocation logic, attention processing, and the user interface.

Backend Enhancements to IP-Adapter:

• New Methods and Logic:

  • Added style_strong and style_precise methods to the IPAdapterInvocation class, with specific logic for determining target attention blocks based on the base type (e.g., sd-1, sdxl). ([[1]](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-c63c2992c76f7ae765f25b13f9ecd2b8bb63eef5e5159dfa218a37926f95f980L97-R98), [[2]](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-c63c2992c76f7ae765f25b13f9ecd2b8bb63eef5e5159dfa218a37926f95f980R151-R182))

• Attention Processing Updates:

  • With the help of matt3o, modified CustomAttnProcessor2_0 to handle negative attention blocks for the style_precise method. This is, admittedly, less sophisticated than processing the entire negative block as a separate point of control, but the amount of work involved in that (and the likelihood of that being needed in the future, given this is for Diffusers/SDXL) seems low. ([invokeai/backend/stable_diffusion/diffusion/custom_atttention.pyR166-R169](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-2674cdb025f9eb53c71d878340376496d4f2cbf9d1ca77d7a048ea031124e8a2R166-R169))
  • Enhanced UNetAttentionPatcher to mark specific blocks as negative when using the style_precise method. ([invokeai/backend/stable_diffusion/diffusion/unet_attention_patcher.pyR43-R50](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-500c7c7bdd4c31b28a0bea8009cd83b12ccb32813cfef4a0db0024607d8a9062R43-R50))

Frontend Updates for IP-Adapter Configuration:

• UI Enhancements:

  • Added style_strong and style_precise options to the IP-Adapter method dropdown in the user interface, along with descriptions. ([invokeai/frontend/web/src/features/controlLayers/components/IPAdapter/IPAdapterMethod.tsxR38-R47](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-a8e1c058bc6b8e47c2b9d696b5fff027d12f45d755977add5ad779ac00254630R38-R47))

• Schema and Type Updates:

  • Extended the zIPMethodV2 and related schemas to include the new methods, ensuring validation and type safety. ([[1]](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-41b99074662523f313cc66e9d6ce8fa9067ac0a87dbf8bcd374554dc533fbcf9L53-R53), [[2]](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-21f94e3b560358a31df37b048dc4077c48c02f38c55068a6ffc4f11be239f07aL150-R150), [[3]](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-7000f59d98ff3b216f5408fecd7c6ff301af310cd1f6d8301d08553fc696d4c4L9173-R9173), [[4]](https://github.com/invoke-ai/InvokeAI/pull/8011/files#diff-7000f59d98ff3b216f5408fecd7c6ff301af310cd1f6d8301d08553fc696d4c4L9299-R9299))## Summary

QA Instructions

Should test all previous instances of IP Adapter vs new
Regression:

• SD1 Style
• SD1 Composition
• SD1 Full
• SDXL Style
• SDXL Composition
• SDXL Full
• Multiple IP Adapters

Functionality:

• SD1 Style Precise
• SD1 Style Strong
• SDXL Style Precise
• SDXL Style Strong

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[hipsterusername]

@psychedelicious -- I'm not entirely sure why the typegen checks are failing -- but i think i've worked with matt3o to get everything wired in correctly, and have tested vs main to confirm no regressions (although as a note, did just test the IP adapter plus models)

[psychedelicious]

Just needed to re-gen the types, looks good now