Apply ConDelta AutoScale

Plain ApplyConDelta has a quirk worth knowing before you use it: strength = 1 doesn't mean the same thing every time. A conditioning's raw vector magnitude depends on your prompt's length and content, and a delta's own magnitude depends on what it was built from - so the same numeric strength can feel strong on one prompt and barely register on another. ApplyConDeltaAutoScale is the fix: it normalizes the delta against your base conditioning's own magnitude before applying strength, so "1" behaves more consistently across different prompts and different saved deltas.

How it works

It's ApplyConDelta plus a normalization step. Before adding the delta, the node computes a magnitude ratio between your base conditioning and the condelta, using whichever statistic ratio_type picks (mean, max, or median), and scales the delta to match before applying strength. There's also an optional hard ceiling: turn on clamp and any value in the scaled delta beyond clamp_value gets capped, which is a cheap insurance policy against a delta that turns out to be much larger than expected once normalized.

The inputs and outputs that matter

• conditioning - your base conditioning.
• condelta - the same file-picker dropdown as plain ApplyConDelta, reading from ComfyUI/models/ConDelta. (Its tooltip also says "The name of the LoRA" - same copy-paste artifact, same non-issue.)
• ratio_type - default median, with mean and max as the alternatives. Median is the sane default; it's less thrown off by a handful of extreme values than mean or max would be.
• strength - default 1, range -100 to 100, applied after the auto-scaling.
• clamp - default off; when on, caps the scaled delta's values.
• clamp_value - default 3, only relevant when clamp is on.

Output is a single CONDITIONING.

How to install it

Via ComfyUI Manager: search "ComfyUI-ConDelta", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/envy-ai/ComfyUI-ConDelta

Then set up the model folder the same way plain ApplyConDelta needs it - create ComfyUI/models/ConDelta (case sensitive) and copy in the pack's sample_models, or your own saved deltas, before restarting.

Common issues & troubleshooting

This is the node to reach for by default, not the plain one. If you're going to be swapping between multiple saved deltas or trying the same delta on prompts of very different lengths, the auto-scaling here saves you from re-tuning strength from scratch every time. Use plain ApplyConDelta when you specifically want raw, unnormalized control.

Empty dropdown means the same thing it always does - no models/ConDelta folder, or an empty one. See the plain ApplyConDelta article for the setup steps if you haven't done them yet.

If you're still seeing burn or artifacts at a strength that used to feel safe elsewhere, turn clamp on. Auto-scaling changes the effective magnitude of the delta, so a strength that was fine unscaled can land somewhere different once normalized - clamping gives you a hard backstop instead of hunting for the right strength by trial and error.

Try mean or max if median isn't giving you the effect you expect, especially on a delta with a very uneven distribution of values - the three ratio types can genuinely disagree on what "normal" looks like for a given delta.