Feature: Add FLUX.2 LOKR model support (detection and loading)

[lstein]

Summary

This is essentially an update to PR #8862, which added Flux.2 Klein LoRA support. Even after this PR was applied, I have found a number of Flux.2 LoRAs out there which work with Comfy, but are not correctly detected and loaded by InvokeAI. They are a LOKR variant.

While most of these are NSFW LoRAs on Civitai, we should probably support them, and this PR fixes that.

Warning: it was mostly written by copilot and needs code review. I've done functional testing.

Related Issues / Discussions

QA Instructions

Here is a SFW LoRA to test with: flux-2-klein-4b-spritesheet-lora. It should download, be recognized as a Flux.2 LoRA and generate properly.

Also check that other flavors of Flux.2 LoRAs still work.

Merge Plan

Simple merge.

Checklist

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

[Pfannkuchensack]

Works fine with the 2 loras i tested.

There are some warning in the console:

[2026-02-25 02:18:11,165]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-double_stream_modulation_img.lin [2026-02-25 02:18:11,166]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-double_stream_modulation_txt.lin [2026-02-25 02:18:11,166]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-final_layer.linear [2026-02-25 02:18:11,166]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-img_in [2026-02-25 02:18:11,341]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-single_stream_modulation.lin [2026-02-25 02:18:11,341]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-time_in.in_layer [2026-02-25 02:18:11,342]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-time_in.out_layer [2026-02-25 02:18:11,342]::[LayerPatcher]::WARNING --> Failed to find module for LoRA layer key: lora_transformer-txt_in

But these messages are also appearing in the flux2 small lora branch.

[Pfannkuchensack]

I just made a commit for the warnings that fixes it.

[lstein]

Thanks!

If the code looks good, could you give it an approval? I cannot do it on my own.

[Pfannkuchensack]

Works fine.

[JPPhoto]

_convert_bfl_layer_patch_to_diffusers() only handles fused QKV when isinstance(layer, LoRALayer), and does not handle LoKR/FullLayer equivalents. That means img_attn.qkv/txt_attn.qkv in a non-LoRALayer patch will fall through and be kept as transformer_blocks.*.img_attn.qkv, which does not exist in diffusers. (invokeai/backend/patches/lora_conversions/flux_bfl_peft_lora_conversion_utils.py:465 and see logic around lines 471–506)

Also, _validate_looks_like_lora() only checks lokr_w1, lokr_w2, hada_w1_a, and hada_w2_a. It doesn't check lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, hada_w1_b, hada_w2_b, hada_t1, and hada_t2. (invokeai/backend/model_manager/configs/lora.py:516 - the suffix list is around lines 540–553)