Apply ConDelta

This is the node the whole pack is built around, and it's the one shown in the README's own screenshot: drop it between your CLIP Text Encode and your KSampler, pick a saved ConDelta from the dropdown, set a strength, and you've applied a style, a concept, or an "anti-concept" to your prompt without touching a LoRA trainer or an embedding trainer. A ConDelta is the vector difference between two encoded prompts - subtract "a city" from "an art deco city" and the leftover vector is roughly "art deco"-ness - and this node is how you add that leftover back into a different prompt's conditioning.

Community reaction on release called it "embedding-lite," which is a fair summary: like an embedding, it's tiny (the pack's own numbers are around 4 MB for Flux, ~0.6 MB for SDXL) and it doesn't touch model weights. Unlike an embedding, it costs literally nothing at generation time - no extra forward pass, no extra download, just a vector add before sampling starts.

How it works

ApplyConDelta loads the ConDelta you pick from disk, and adds it to your input conditioning, scaled by strength. That's it - no CFG involvement, no extra compute, because the addition happens once on the conditioning tensor before the sampler ever runs.

The inputs and outputs that matter

• conditioning - the conditioning you're modifying, normally straight out of a CLIP Text Encode.
• condelta - a dropdown populated from your ComfyUI/models/ConDelta folder. Its tooltip says "The name of the LoRA," which is a copy-paste leftover in the author's code - this node has nothing to do with LoRAs, ignore it.
• strength - default 1, range -100 to 100. Positive pushes toward the concept the delta encodes; negative pushes away from it, which is how this pack doubles as a negative-prompt alternative.

Output is a single CONDITIONING, ready for your sampler.

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, critically, set up the model folder - this is the step people miss and the dropdown will be empty without it:

1. Create a folder named ConDelta (case sensitive) inside ComfyUI/models.
2. Copy the files from custom_nodes/ComfyUI-ConDelta/sample_models into that new folder.
3. Restart ComfyUI so the dropdown picks them up.

Common issues & troubleshooting

Empty dropdown. If condelta shows no options, you skipped the folder step above. No models/ConDelta folder (or an empty one) means nothing to select.

Match the delta to your model family. The README's own warning is blunt: "make sure to select a ConDelta that's compatible with the model you're using." A delta computed from Flux's T5-based conditioning isn't meaningful applied to an SDXL CLIP conditioning - different vector space entirely.

Strength too high burns the image. Same failure mode as pushing CFG or a LoRA weight too far - structure breaks down, colors oversaturate. Start near the default and creep outward; the useful range is usually much smaller than the ±100 the slider allows.

Results can be inconsistent. This is a genuinely mixed-reception node in the community - some people get exactly what they expect (a Flux-compatible negative-prompt substitute, a lightweight style push), others report it mostly reproducing the source prompt with extra bleeding, especially on SDXL. Treat it as a cheap thing to experiment with, not a guaranteed LoRA replacement.