Krea2 Projector Delta Advance

Krea2 Projector Delta Advance is the same idea as its plain sibling - patch Krea 2's diffusion_model.txtfusion.projector.weight with a twelve-value delta to clear the baked-in refusal filter - but instead of one comma-separated text box you get twelve labelled number inputs, and it hands back a couple of extra outputs you'll actually want. It's the same pack, same install, same one weight; the difference is ergonomics and bookkeeping.

This is the node to reach for when you're tired of squinting at a string like 0,0,0,0,0,0,0,0,-0.51171875,-0.890625,0,0 and wondering which column is which. Krea 2's filter is a (1,12) delta on the text-fusion projector - the community analysis that this pack is based on showed FilterBypass v2 only touches columns 9–10, v3 adds column 11, and the LoRA-style bypass hits all twelve. The Advance node puts every column on screen as its own input with fine precision (step 0.0001), which is the difference between guessing and seeing.

How it works

Mechanically it's the same model patch: the node clones your model and applies the twelve values to the projector weight via ComfyUI's add_patches with a LoRA-style strength multiplier. What's different is how values get in. The d1–d12 slots are the editor. Pick a preset - FB2, FB3, FEDOR (numerically identical to FB2), or SKC3VO (all twelve columns) - and the JS extension fills the slots for you; the slots are the custom values, so there's no separate text box here. When use_custom is on, the preset is ignored and whatever's in the slots is used.

Inputs and outputs

The inputs worth touching: preset, use_custom, strength (default 1, −100 to 100, scales the whole vector), and the d1–d12 floats (each −1000 to 1000). That's the whole list - nothing hidden.

Outputs, and this is the part the plain node doesn't give you:

• model - the patched model, into your KSampler as usual.
• delta_text - the twelve values actually applied, after strength scaling, formatted as a comma string. This is gold for sharing: copy it into the plain Krea2 Projector Delta node, or save it as a recipe.
• preset_name - custom or the preset you picked, handy when you're logging runs and comparing.

After each run the applied values write back into the slots, so what you see on the node is what the model actually got - no mystery about which strength made that image.

Installing

Same pack as the plain node, same steps:

cd ComfyUI/custom_nodes
git clone https://github.com/Patvessel/ComfyUI-krea2_projector_delta

Restart ComfyUI and hard-refresh the browser so the JS extension loads. No pip install - the pack has no dependencies beyond what stock ComfyUI ships. It only does anything on an actual Krea 2 model; on any other checkpoint there's no txtfusion.projector.weight to patch and the node is inert, so double-check your model loader.

Gotchas

• Slot values vs. preset values. With use_custom off, the d-slots are a mirror of the preset - edit them all you like, they won't stick once a run happens. Flip use_custom on before treating them as your own.
• Precision bites both ways. The 0.0001 step is great for fine deltas and terrible for fat-fingering - a stray 1 where you meant 0.1 on column 9 is a very different image. This is where the fixed-seed A/B test matters.
• Restart after editing presets.json to rebuild the dropdown, and refresh the browser if the auto-fill stops behaving; frontend and backend load separately.

Which of the two nodes should you pick? The plain one if you think in comma strings and want the single text box. The Advance if you want every column visible, tunable in isolation, and a written record of what actually ran - which, once you start comparing FB2 against SKC3VO at different strengths, is most people.