Skip to content

Comments

Feature: flux2 klein lora support#8862

Merged
lstein merged 10 commits intoinvoke-ai:mainfrom
Pfannkuchensack:feature/flux2-klein-lora-support
Feb 24, 2026
Merged

Feature: flux2 klein lora support#8862
lstein merged 10 commits intoinvoke-ai:mainfrom
Pfannkuchensack:feature/flux2-klein-lora-support

Conversation

@Pfannkuchensack
Copy link
Collaborator

@Pfannkuchensack Pfannkuchensack commented Feb 8, 2026

Summary

Add full LoRA support for FLUX.2 Klein models (4B and 9B variants), including automatic detection, loading, key conversion, and UI integration.

Model Detection & Configuration

  • Detect FLUX.2 Klein LoRAs during model probe by inspecting tensor shapes (context_embedder, vector_in, hidden size, MLP ratio) to distinguish them from FLUX.1 LoRAs
  • Auto-detect Klein variant (4B vs 9B) from tensor dimensions and store as variant field on the config
  • New LoRA_LyCORIS_Flux2_Config and LoRA_Diffusers_Flux2_Config model configs
  • Register FluxLoRAFormat.BflPeft as a new recognized LoRA format
  • Fix Config_Base.get_tag() to handle None defaults (needed for the optional variant field)

BFL PEFT LoRA Format Support

  • New flux_bfl_peft_lora_conversion_utils.py with full BFL PEFT format detection and conversion
  • Handles fused QKV splitting (BFL fuses Q/K/V, diffusers uses separate projections)
  • Two conversion paths: FLUX.1 (keeps BFL keys) and FLUX.2 Klein (converts to diffusers keys)
  • Runtime fallback conversion via convert_bfl_lora_patch_to_diffusers() for LoRAs loaded with FLUX.1 format but applied to FLUX.2 models

LoRA Loading & Application

  • New Flux2KleinLoRALoaderInvocation and Flux2KleinLoRACollectionLoader invocations for applying LoRAs to Klein's transformer and Qwen3 text encoder
  • Variant mismatch warnings when a LoRA's variant doesn't match the transformer

Frontend

  • New addFlux2KleinLoRAs.ts graph builder for wiring LoRAs into the Klein generation graph
  • LoRA picker filters by variant compatibility (only shows 4B LoRAs for 4B models, etc.)

Related Issues / Discussions

QA Instructions

  1. Install a FLUX.2 Klein model (4B or 9B)
  2. Install a LoRA trained for that Klein variant (in BFL PEFT, diffusers, or LyCORIS format)
  3. Verify the LoRA is auto-detected as flux2 base with the correct variant (4B/9B) in the Model Manager
  4. Verify the LoRA picker only shows variant-compatible LoRAs when a Klein model is selected
  5. Generate an image with the LoRA applied and verify it takes effect without shape errors
  6. If possible, test with a mismatched variant LoRA and verify a warning is logged

Merge Plan

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 and others added 3 commits February 6, 2026 01:01
@Pfannkuchensack @claude
WIP: Add FLUX.2 Klein LoRA support (BFL PEFT format)
f5d528f 
Initial implementation for loading and applying LoRA models trained
with BFL's PEFT format for FLUX.2 Klein transformers.

Changes:
- Add LoRA_Diffusers_Flux2_Config and LoRA_LyCORIS_Flux2_Config
- Add BflPeft format to FluxLoRAFormat taxonomy
- Add flux_bfl_peft_lora_conversion_utils for weight conversion
- Add Flux2KleinLoraLoaderInvocation node

Status: Work in progress - not yet fully tested

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@Pfannkuchensack
feat(flux2): add LoRA support for FLUX.2 Klein models
e1963e9 
Add BFL PEFT LoRA support for FLUX.2 Klein, including runtime conversion
of BFL-format keys to diffusers format with fused QKV splitting, improved
detection of Klein 4B LoRAs via MLP ratio check, and frontend graph wiring.
@Pfannkuchensack
feat(flux2): detect Klein LoRA variant (4B/9B) and filter by compatib…
3fa6ebe 
…ility

Auto-detect FLUX.2 Klein LoRA variant from tensor dimensions during model
probe, warn on variant mismatch at load time, and filter the LoRA picker
to only show variant-compatible LoRAs.
@github-actions github-actions bot added python PRs that change python files invocations PRs that change invocations backend PRs that change backend files frontend PRs that change frontend files labels Feb 8, 2026
@Pfannkuchensack
Copy link
Collaborator Author

Pfannkuchensack commented Feb 9, 2026

I added the variant Fields for Loras too and maybe we could use that for SDXL Subtypes too.

@Pfannkuchensack Pfannkuchensack marked this pull request as ready for review February 9, 2026 13:57
@lstein lstein self-assigned this Feb 16, 2026
@lstein
Copy link
Collaborator

lstein commented Feb 17, 2026

It isn't clear to me how to find BFL PEFT or diffusers LoRAs for Flux.2. Everything I found was a .safetensors fle. If you could provide some guidance for how to distinguish the formats, I can test more comprehensively.

In the meantime, here are some popular Flux.2 LoRAs I tried:

@Pfannkuchensack
Copy link
Collaborator Author

Pfannkuchensack commented Feb 17, 2026

I take a Look at the ones that are not working.

@lstein lstein added the v6.12.0 Intended for 6.12.0 release label Feb 20, 2026
@digitalassassins
Copy link

digitalassassins commented Feb 22, 2026

Good job on this, was trying to get bfs_head_v1_flux-2-klein_9b Lora working in 6.11 and just wasn't having anything, will wait for 6.12 😁

@Pfannkuchensack
Merge branch 'main' into feature/flux2-klein-lora-support
ba39cf9
@Pfannkuchensack
Fix detection and loading of 3 unrecognized Flux.2 Klein LoRA formats
35d83c1 
Three Flux.2 Klein LoRAs were either unrecognized or misclassified due to
format detection gaps:

1. PEFT-wrapped BFL format (base_model.model.* prefix) was not recognized
   because the detector only accepted the diffusion_model.* prefix.
2. Klein 4B LoRAs with hidden_size=3072 were misidentified as Flux.1 due to
   a break statement exiting the detection loop before txt_in/vector_in
   dimensions could be checked.
3. Flux2 native diffusers format (to_qkv_mlp_proj, ff.linear_in) was not
   detected because the detector only checked for Flux.1 diffusers keys.

Also handles mixed PEFT/standard LoRA suffix formats within the same file.
@lstein
Copy link
Collaborator

lstein commented Feb 24, 2026

The Spritesheet is working. However, the Spritesheet comfy converted version is still unrecognized.

Consistence Edit LoRA is now recognized and works.

I think this covers enough of the use cases for now and will Accept it.

lstein
lstein approved these changes Feb 24, 2026
View reviewed changes
Copy link
Collaborator

@lstein lstein left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Works as advertised.

@lstein lstein enabled auto-merge (squash) February 24, 2026 02:51
@lstein lstein merged commit afbd45a into invoke-ai:main Feb 24, 2026
13 checks passed
@Pfannkuchensack Pfannkuchensack deleted the feature/flux2-klein-lora-support branch February 24, 2026 13:41
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@lstein lstein lstein approved these changes

@blessedcoolant blessedcoolant Awaiting requested review from blessedcoolant blessedcoolant is a code owner

@dunkeroni dunkeroni Awaiting requested review from dunkeroni dunkeroni is a code owner

@JPPhoto JPPhoto Awaiting requested review from JPPhoto JPPhoto is a code owner

Assignees

@lstein lstein

Labels

backend PRs that change backend files frontend PRs that change frontend files invocations PRs that change invocations python PRs that change python files v6.12.0 Intended for 6.12.0 release

Projects

Invoke - Community Roadmap
Status: 6.12.x

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@Pfannkuchensack @lstein @digitalassassins