Conditioning
The encoded prompt, with any hints attached, that steers the sampler.
In plain words
Conditioning is what comes out of a text node: your prompt turned into numbers by the text encoder. The sampler cannot read text; it reads conditioning. Other nodes can add hints to it on the way: a ControlNet picture, an area of the frame, a reference image, the guidance number of Flux, the start frame of a video.
Every sampler has two conditioning inputs:
positive: what you want. The model moves towards it.negative: what you do not want. The model moves away from it (only when CFG is above 1, see CFG).
Where you meet it in a graph
- CLIP Text Encode (Prompt) makes it:
CLIPand text in,CONDITIONINGout. - Apply ControlNet takes positive and negative and gives them back with the control hint added.
- FluxGuidance writes the guidance number into the conditioning for Flux dev.
- Conditioning (Combine) mixes two conditionings; Conditioning (Concat) joins them into one longer one; Conditioning (Set Area) limits one to a rectangle of the picture.
- Video and inpainting start nodes (WanImageToVideo, InpaintModelConditioning) take positive and negative and give them back with the picture attached. From there on the sampler must take the conditioning from their outputs, not straight from the text nodes.
Typical mistakes
- positive and negative swapped: the picture avoids what you described.
- A sampler fed from the text nodes directly in a video from a picture graph: the picture never reaches the model, the clip is drawn from nothing.
- Conditioning from one model's encoder going to another model's sampler: it means nothing to the second model.
- Empty positive prompt: the model draws something random.
How the app helps
Select a CONDITIONING link and the Explain tab says where it comes from and where it should go. Check warns about an empty positive prompt and about a negative prompt that does nothing at cfg 1.