CLIP Text Encode (Advanced)

This is the pack's "I understand your fancy new model" node. CLIP Text Encode (Advanced) [Eclipse] takes a plain prompt, encodes it, and then gives you two dials the stock encoder doesn't have: a global multiplier for overall prompt strength, and per-layer rebalancing for Krea 2's multi-layer tap architecture. If you're not running Krea 2, you still get a handy prompt-strength multiplier out of it.

Here's the mechanism, grounded in the source. A CLIP embedding for a model like Krea 2 isn't one monolithic vector - the feature dimension is split into 12 layer taps, each carrying different semantic weight. When you pick a rebalance_preset, the node multiplies each tap by a fixed gain vector:

• balanced - boosts the late taps up to 5x (that's the [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 2.5, 5.0, 1.1, 4.0, 1.0] preset).
• detail - dampens early taps, hammers late ones up to 6x. More aggressive, for when you want fine detail emphasized.
• subtle - gentle late-tap boost, up to 2x.
• uniform - a flat 1.0 across the board, i.e. no rebalancing.
• custom - you supply per_layer_weights as 12 comma-separated floats.

Then renormalize (on by default) rescales the altered tensor's RMS back to the original, so you get the rebalancing without accidentally cranking the overall prompt volume. The multiplier is the separate global knob on top: >1.0 amplifies the whole conditioning, <1.0 weakens it. And krea2_only_multiplier is the polite option - the multiplier applies only when the model actually is Krea 2, so you can leave a multiplier in the graph without it wrecking a normal SDXL workflow.

For non-Krea 2 models the node is smart about it: if the conditioning dimension isn't divisible by your weight count, it logs a warning and falls back to a uniform scale, so the layer math never silently corrupts a normal model. The multiplier still applies.

If this rebalancing concept sounds familiar, that's because it is: Krea 2 shipped the same day a dedicated ComfyUI-ConditioningKrea2Rebalance node appeared in the community, and it quickly became the standard way to shape Krea 2 output. This encoder is this pack's in-graph version of that idea - same adjustment, no extra pack.

Inputs

• clip and text - as any encoder.
• rebalance_preset - none / balanced / detail / subtle / uniform / custom.
• per_layer_weights - active only on custom; 12 floats for Krea2, commas or semicolons accepted.
• multiplier, renormalize, krea2_only_multiplier - described above.

Output: one conditioning socket, same as the basic encoder.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_SmartModelLoader.git
cd ComfyUI_SmartModelLoader
python -m pip install -r requirements.txt

or ComfyUI Manager → ComfyUI Smart Model Loader, restart.

Where people trip

• "per_layer_weights is empty" - you set custom but left the field blank. It needs at least 2 numbers; for Krea2, 12.
• Everything looks identical to stock - with preset none and multiplier 1.0, that's exactly what it is. This node earns its keep only when you use its knobs.
• Overdone detail - detail at 6x on late taps is a lot. If output gets crunchy or text-heavy artifacts appear, back off to balanced or subtle before blaming the model.