DIT Deep Debiaser (FLUX.2 Klein — Verified)

Klein's known weakness is anatomy - extra fingers, a third arm, mutations that get baked in during the first couple of denoising steps and then compound. Block-level LoRA editing doesn't reach that, because the problem lives in the base model's own architecture, not in a LoRA. This node goes one level deeper: instead of toggling whole blocks, it patches individual attention heads, MLPs, and norms inside them.

What "debiasing" means here, and why it goes past block editing

Klein's block-level control (the pack's own selective LoRA loaders) works on a LoRA's contribution to 25 or 57 blocks depending on model size. This node works on the base checkpoint's verified architecture directly - 63 individually addressable sub-components across Klein's 8 double blocks and 24 single blocks, plus 7 global components (img_in, txt_in, time_in, per-modulation controls, final_layer). The pack states this architecture was verified through forward-pass hook tracing, not just inferred from tensor names, which matters because Klein's double and single blocks behave genuinely differently: double blocks (0-7) keep image and text tokens in separate streams with no cross-modal interaction, while single blocks (0-23) concatenate everything into one 8,652-token sequence - which the node's own description names directly as "where the model overwrites your image" when generation goes wrong.

The pack's own author found this useful on release week for exactly the anatomy problem: a Reddit post from the account behind the pack demonstrated that a handful of early single-block joint components (sb0 through sb4, plus sb9) - with sb2 singled out as the highest-leverage - control how much detail preservation happens on an edit, independent of the other 57 unchanged components. Lowering sb2's strength was reported as the main lever for keeping original detail rather than letting Klein regenerate it.

Inputs and outputs that matter

• model - a Flux 2 Klein checkpoint (4B or 9B; the verified mapping applies to the Klein architecture generally).
• preset - a large set of pre-built weakening presets (Weaken All Singles 90/85%, Weaken Late Singles, Weaken DB img_mlp/txt_mlp/img_attn/txt_attn 90%, and more) that give you a sane starting point instead of hand-tuning 63 sliders cold.
• img_in, txt_in, time_in, db_mod_img, db_mod_txt, sb_mod, final_layer - the 7 global controls.
• db0_img_attn/db0_img_mlp/db0_txt_attn/db0_txt_mlp through db7_* - per-double-block controls, 4 sub-components × 8 blocks = 32.
• sb0 through sb23 - one control per single block, 24 total. Each of the 63 controls has its own toggle and a -5 to 5 strength slider.
• save_model / save_mode / save_filename / save_directory (optional) - export the modified model. save_mode gives you full_model (the complete ~18GB checkpoint, load directly as a replacement) or diff_only (just the changes, applied as an additive patch elsewhere).
• Outputs: model (patched, LoRA-safe - the pack applies changes via ComfyUI's add_patches system, so LoRAs still apply correctly on top), info (a text report of what changed).

Installing it

ComfyUI Manager: search "Realtime LoRA Trainer." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ShootTheSound/comfyUI-Realtime-Lora

Restart ComfyUI. The README is explicit that this whole "Deep Debiaser & Inspector" suite works out of the box - no additional installation beyond the pack itself, same as the analyzers and loaders. You do need a Flux 2 Klein checkpoint already downloaded and working, which is a separate, much larger requirement handled entirely outside this node.

Troubleshooting

Set a sub-component to 0.0 and the output turned to noise. This is explicitly called out - the pack's own warning is that any sub-component at 0.0 will likely produce noise or artifacts, and the recommended working range is 0.5-1.5. Zero isn't "off," it's closer to "removed the load-bearing wall."

Trying to fix identity drift on an edit and nothing here helped. This node targets the DiT (the diffusion transformer). Klein's known color-shift-on-edit problem lives partly in the VAE, and text-encoder-side prompt issues live in the Qwen3 encoder - both have their own separate debiaser nodes in this pack. Match the symptom to the right component before assuming the DiT debiaser should fix everything.

Saved a diff_only file and it doesn't load as a standalone model. That's by design - diff_only is meant to be applied as an additive patch on top of the original checkpoint, not loaded on its own. Use full_model if you want something that works as a drop-in replacement checkpoint.