Skip to content

ComfyUI Flux2Klein Enhancer

Intermediate

ComfyUI-Flux2Klein-Enhancer is a third-party ComfyUI custom-node package for reference conditioning and identity/detail transfer. It targets FLUX.2 Klein 9B behavior and is not an official FLUX model release. Status verified against the project repository on 2026-08-27.

Current Status

Property Verified state
Repository capitan01R/ComfyUI-Flux2Klein-Enhancer
Verified default-branch revision 6804643bff9a20926106427ff08d5b1bd2e49861
License MIT
Primary target FLUX.2 Klein 9B
Reference count Up to eight references in the current workflow
Neutral controls Exact neutral values are intended as pass-through
Independent identity benchmark Not available in the bounded research pass

The package grew beyond the two-node form described in the initial news item. Current workflow concepts include reference encoding, Multi ReferenceLatent, Identity Feature Transfer Final, and optional masks.

Development History

Date Event Temporal status
2026-04-07 Initial enhancer package/report Historical foundation
2026-06-24 Expanded workflow/package reported Current successor event
2026-07-11 Verified repository revision date Current repository snapshot used here

Keep the original news text unchanged. Attach later package expansion and current repository state as successor nodes.

Installation and Pinning

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer.git
cd ComfyUI-Flux2Klein-Enhancer
git checkout 6804643bff9a20926106427ff08d5b1bd2e49861
git rev-parse HEAD

Restart ComfyUI and load a workflow that matches the pinned revision. Preserve the workflow JSON with the run receipt; node names and sockets may drift across revisions.

Evaluation Protocol

Run a same-seed comparison:

comfyui_revision: <exact-revision>
plugin_revision: 6804643bff9a20926106427ff08d5b1bd2e49861
model: <exact-flux2-klein-artifact>
reference_count: <1-to-8>
reference_order: <ordered-list>
masks: <file-hashes-or-none>
seed: <integer>
dimensions: <width>x<height>
steps: <integer>
enhancer_controls: <all-values>
gpu_and_vram: <exact-device>

Compare a baseline workflow and an enhancer workflow while holding model, prompt, seed, steps, dimensions, and sampler fixed. Record identity similarity, unwanted detail transfer, prompt adherence, and peak VRAM. Neutral settings should be tested as an exact pass-through invariant.

Evidence Boundaries

  • The repository proves implementation identity, node surface, target model, and license.
  • “No LoRA needed,” exact identity preservation, quality gains, and low-VRAM behavior are author/community claims until independently measured.
  • Smaller conditioning models may be recognized by the implementation, but the hook schedules and documented target center on Klein 9B.
  • No qualified Chinese community source was retained in the initial bounded pass.

Community and Issue Evidence

  • Issue #35 found that Identity Feature Transfer Final returned NaN from IS_CHANGED, forcing fixed-seed downstream reruns. The owner merged a fix on 2026-06-12. This is a historical bug; revisions after that merge should not inherit the workaround.
  • Issue #41 reports identity changes after ComfyUI v0.28.0 with two masked references. The reporter isolated zero_unmasked_tokens and used focus_only as a workaround; the issue remained open. Preserve ComfyUI version, mask wiring, preset, and plugin revision when reproducing.
  • The exact community showcase thread is useful for discovering the method, but “exact preservation” remains an author/community claim without an independent metric.

Gotchas

  • Issue: Treating the plugin as an official FLUX feature -> Fix: identify it as a third-party ComfyUI package and pin its commit.
  • Issue: Claiming identity preservation from a showcase -> Fix: run same-seed baseline/enhancer comparisons with an explicit review metric.
  • Issue: Loading an old two-node workflow against the expanded package -> Fix: match workflow JSON to the exact plugin revision.
  • Issue: Assuming smaller conditioning support equals full 9B parity -> Fix: verify the exact model path and schedules locally.
  • Issue: Reusing zero_unmasked_tokens across ComfyUI versions -> Fix: reproduce with the pinned mask/preset setup; focus_only is a reported v0.28 workaround, not a universal default.

Temporal Status

  • Current: expanded multi-reference workflow at the pinned July 2026 revision.
  • Historical: initial April two-node description.
  • Reported, not independently verified: exact preservation, no-LoRA replacement, quality gain, and low-memory performance.

Agent Brief

Pin the plugin, ComfyUI, workflow JSON, and model before troubleshooting. Separate repository facts from showcase/community efficacy claims. Require a controlled baseline for identity or detail claims. Preserve the initial news as history and attach current package state as the endpoint rather than rewriting it.

Sources

  • Project repository: https://github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer
  • Community discussion lane: https://www.reddit.com/r/StableDiffusion/comments/1t1mvyh/flux2klein_exact_preservation_no_lora_needed/
  • Fixed cache-invalidation bug: https://github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer/issues/35
  • Open ComfyUI v0.28 mask behavior: https://github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer/issues/41