Text Encoder Deep Debiaser (Qwen3-4B)

Z-Image and Flux 2 Klein 4B both hand your prompt to a Qwen3-4B language model instead of CLIP, and that model isn't a flat bag of tokens - it's 36 transformer layers, each doing something a little different to your prompt before it ever reaches the diffusion model. This node lets you turn individual pieces of those layers up or down, rather than accepting the encoder's behavior as fixed.

What it is and why you'd reach for it

Where the pack's LoRA loaders work on blocks of the diffusion model, this works on the text encoder - the component that turns your written prompt into the conditioning vectors everything downstream reacts to. Because Qwen3-4B is a proper language model, your prompt isn't read as a bag of CLIP-style tags; it's wrapped in a chat template and processed layer by layer like an instruction. That means a layer that's over-weighting one token, or under-attending to spatial relationships you wrote, is a real, addressable thing rather than folklore - and this node is the tool for addressing it directly instead of fighting it with prompt phrasing tricks.

It's one of five Deep Debiaser nodes in ShootTheSound/comfyUI-Realtime-Lora's newer sub-component suite, alongside a matching Qwen3-8B version (for Klein 9B) and DiT/VAE debiasers for the diffusion side.

How it works

Qwen3-4B's 36 layers each split into five addressable parts: input_norm, attn (self-attention), attn_norm, mlp (the feed-forward block), and post_norm - 180 controls total, plus a single embed_tokens control and a final_norm, for 182 sub-components. Each gets its own strength multiplier, 1.0 unchanged, working below and above that to weaken or boost that specific piece.

The one quantization caveat worth knowing before you start: this only modifies float tensors - norms and some weights. Quantized layers, if your text encoder is loaded in a quantized format (U8, F8_E4M3), are preserved unchanged rather than patched, because you can't cleanly scale a quantized weight the same way. If a strength change on a quantized-heavy layer does nothing, that's why.

Inputs and outputs that matter

• clip - the Qwen3-4B text encoder to edit (used by Z-Image and Klein 4B).
• preset - 13 built-in options (Weaken ALL attn 90/85/80%, Weaken ALL mlp, Weaken ALL attn+mlp, Weaken ALL norms, Global 95%/90%, and more) rather than hand-setting 182 sliders from a blank state.
• embed_tokens, l0_input_norm through l35_post_norm (the full 36-layer × 5-part grid), and final_norm - each with a toggle and a -5 to 5 strength slider.
• save_model / save_mode / filename (optional) - export the modified encoder. save_mode offers full_model, diff_only, or both.
• Outputs: clip (patched, feed to your text-encode node), 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 further install needed - the Deep Debiaser suite runs out of the box, same as the rest of the pack's non-training nodes. You need a Qwen3-4B text encoder already loaded, which Z-Image and Klein 4B workflows already require regardless of this node.

Troubleshooting

Set a sub-component to 0.0 and prompt following broke entirely. Same warning as the pack's other debiasers: 0.0 tends to produce noise or a broken result rather than a clean absence, and the sane working range is 0.5-1.5. Zero is a stress test, not a normal setting.

Boosted embed_tokens to fix weak prompt adherence and nothing changed. Check whether your text encoder is loaded quantized - embed_tokens is float and should be modifiable regardless, but if the layer you actually needed to touch is one of the quantized attention or MLP weights, the change silently doesn't apply. The node's own tip is that boosting embed_tokens to 1.5-2.0 helps prompt adherence, which is worth trying first since it's the one control least likely to be quantization-locked.

Trying to fix an anatomy or color problem and nothing here helps. This node only touches the text encoder - it changes how your prompt gets interpreted, not how the diffusion model renders or how the VAE decodes. Anatomy issues on Klein live in the DiT; color-shift issues live in the VAE. Both have their own dedicated debiaser nodes in this pack.