DIT Deep Debiaser (Z-Image Sub-Component)

The pack's Z-Image Model Layer Editor lets you scale whole layers of the base model. This node goes further: inside each of those layers, attention, feed-forward, and normalization are separate, individually addressable pieces - 174 of them across the whole DiT - for when a block-level edit is too blunt an instrument.

What it is and why you'd reach for it

Z-Image's DiT is built from a context_refiner, 30 main transformer layers, and a noise_refiner, and this node splits every one of them into functional sub-components rather than treating a layer as one knob. Main layers get five: adaLN (timestep and conditioning modulation), attn (self-attention: qkv, output projection, norms), attn_norm, ffn (the SwiGLU feed-forward network), and ffn_norm. The context_refiner and noise_refiner blocks get four each (no adaLN). Add five embedders (cap_embedder, t_embedder, x_embedder, cap_pad_token, x_pad_token) and one final_layer, and you land at 174 total controls.

This is the deepest edit available in the pack for Z-Image's diffusion transformer - deeper than the block-level Model Layer Editor, and operating on the base checkpoint rather than a LoRA, so whatever you do here applies to every LoRA and every prompt afterward. The pack's own author has used this level of granularity productively in public: block-level community testing on this same pack found an anatomical feature concentrated in a narrow layer band (17-21) rather than where the trainer expected, and sub-component control is the natural next step once block-level analysis has told you roughly where to look and you need finer resolution than "the whole layer."

Inputs and outputs that matter

• model - the Z-Image checkpoint to edit.
• preset - 15 built-in options (Weaken ALL attn 90/85/80%, Weaken ALL ffn 90/85%, Weaken ALL adaLN 90/85%, Weaken ALL attn+ffn, Weaken ALL attn_norm+ffn_norm 90%, and more) as a starting point instead of hand-setting 174 sliders.
• cap_embedder, t_embedder, x_embedder, cap_pad_token, x_pad_token - the 5 embedder controls.
• cr0_attn/cr0_attn_norm/cr0_ffn/cr0_ffn_norm, cr1_* - the context_refiner's two blocks, 4 controls each.
• l0_adaLN/l0_attn/l0_attn_norm/l0_ffn/l0_ffn_norm through l29_* - the 30 main layers, 5 controls each (150 total).
• nr0_* / nr1_* - the noise_refiner's two blocks, 5 controls each (10 total, since these do have adaLN).
• final_layer - the last control.
• Every one of the 174 has its own toggle and a -5 to 5 strength slider.
• save_model / save_mode / filename (optional) - export the modified checkpoint. save_mode gives full_model (a complete, loadable checkpoint), diff_only (just the changes, applied as a patch), or both.
• Outputs: model (patched, feed to your sampler or a LoRA loader downstream), 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. No additional install needed - the Deep Debiaser suite works out of the box, same as the pack's loaders and analyzers, and Z-Image's smaller size (6B parameters, quantizable down to 4GB) means this is one of the more approachable models in the pack to experiment on without a huge VRAM budget.

Troubleshooting

Set a sub-component to 0.0 for a clean removal and got noise. Expected - the pack's own warning applies here too: 0.0 tends to produce noise or artifacts rather than "no contribution," and the recommended working range is 0.5-1.5.

Not sure where to even start among 174 controls. Run the pack's block-level Model Layer Editor or one of its LoRA analyzers first to narrow down which of the 30 main layers is actually relevant to your problem, then come here and use this node's sub-component controls only on that narrower range instead of guessing across all 174 cold.

Edited ffn on several layers hoping to fix a style leak and likeness broke instead. The feed-forward network and attention do different jobs - attention tends to carry more of the relational/positional information (what's next to what), while the feed-forward network processes each token's features independently. If likeness degraded, you may have hit a layer where ffn was actually load-bearing for identity, not just style; walk the change back to a single layer at a time before applying it broadly.