Text Encoder Deep Debiaser (Qwen3-8B / Flux 2 Klein)

What it is

This is the deepest control this pack offers over how FLUX Klein 9B reads your prompt - not per-layer, but per functional sub-component inside each of the encoder's 36 layers: the input norm, the attention block, the attention's own norm, the MLP, and the post-norm, each independently scalable. That's 5 controls x 36 layers, plus embed_tokens and a final_norm, for 182 total dials. Klein's text encoder is Qwen3-8B, a genuine language model reading your prompt as an instruction, and the community has already documented real, specific problems that live at exactly this level of granularity - identity drift on edits, and a quality gap the community itself argues about between the fp8 and full-precision versions of this same encoder. This node is for chasing problems that specific: not "the LoRA is too strong," but "this one attention block is doing something I don't want."

Be honest with yourself about when you need this. It's a scalpel for a narrow class of problem - most prompt issues on Klein are fixed by writing plainer, more direct sentences (Klein reads instructions, not tag soup) long before you need to touch a sub-component weight.

How it works

Every sub-component gets a strength multiplier: 1.0 is unchanged, below 1.0 weakens it, above 1.0 boosts it. It patches through ComfyUI's own add_patches system, which the pack notes is LoRA-safe - your LoRAs still apply correctly stacked on top of whatever you change here. One real constraint worth knowing: the underlying model ships partly quantized, and only the float tensors (norms, some weights) are actually modifiable; the quantized layers (U8, F8_E4M3) stay untouched regardless of what you set here.

The recommended workflow, per the pack itself: run Text Encoder Inspector (Qwen3-8B / Flux 2 Klein) first to find which sub-components actually respond to your specific prompt, then come here and adjust just those, rather than guessing across 182 controls blind.

Inputs and outputs that matter

• clip - the Qwen3-8B encoder.
• preset - quick starting points like "Weaken ALL attn 90%," "Weaken ALL mlp 85%," "Global 95%," and more, alongside Custom and Default.
• embed_tokens / embed_tokens_str and final_norm / final_norm_str - the two global controls; the pack's own tip is that boosting embed_tokens to 1.5–2.0 tends to improve prompt adherence.
• l0_input_norm through l35_post_norm (five sub-components per layer, 36 layers) - every one individually toggleable with its own strength slider.
• save_model, save_mode (full_model, diff_only, or both), filename - export your modified weights to disk instead of applying them live every run.

Outputs: clip (the patched encoder) and info (a summary of what got modified).

Installing it

No external backend required - this node works the moment the pack is installed. ComfyUI Manager (search "Realtime LoRA Trainer") or:

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

Restart ComfyUI.

Common issues

Setting a strength to 0.0 produces noise or broken output. Expected - the pack warns about this directly. Stick to roughly 0.5–1.5 for real adjustments; 0.0 is a diagnostic extreme, not a usable setting.

Nothing changes despite adjusting a sub-component. Check whether you're touching a quantized layer - U8 and F8_E4M3 tensors in this model are preserved unchanged regardless of the slider, by design, since only float tensors are actually patchable.

182 controls is too many to reason about without a starting point. This is exactly what the Inspector node is for - run it first, read which sub-components actually activated for your prompt, and adjust only those rather than sweeping through every layer.

Trying to use this to fix an image-quality problem, not a text-comprehension one. This node reshapes how the encoder reads and weighs your prompt - it has no direct control over the diffusion model or the VAE. If the problem is in generated pixels rather than prompt following, look at FLUX Model Layer Editor or VAE Deep Debiaser territory instead.