Data types
MODEL, CLIP, VAE, LATENT, IMAGE, CONDITIONING and the other kinds of data a link carries.
In plain words
Every link carries one kind of data, its type. The type decides which input can take the link. This page lists the types a newcomer meets in built-in nodes. Packs add their own.
The common types
| Type | What it carries | Comes from | Goes to |
|---|---|---|---|
MODEL | The diffusion model, the part that removes noise and draws | Load Checkpoint, Load Diffusion Model, LoRA loaders, model tuning nodes | Samplers, LoRA loaders, ModelSampling nodes |
CLIP | The text encoder, the part that reads the prompt | Load Checkpoint, Load CLIP, DualCLIPLoader, Load LoRA | CLIP Text Encode and other text nodes |
VAE | The model that turns pixels into a latent and back | Load Checkpoint, Load VAE | VAE Decode, VAE Encode, inpainting and video start nodes |
CONDITIONING | An encoded prompt plus any hints attached to it (ControlNet, area, guidance) | Text nodes, ControlNet, guidance and combine nodes | positive and negative inputs of samplers |
LATENT | A picture or video in compressed form; a video latent is compressed in time too | Empty latent nodes (a video one has a length), VAE Encode, video start nodes, samplers | Samplers, latent upscale, VAE Decode |
IMAGE | Ordinary pictures in pixels, always a batch (one or more) | Load Image, VAE Decode, image nodes | Save Image, Preview Image, VAE Encode, upscalers, ControlNet |
MASK | A black and white map of what to change | Load Image, mask nodes, the mask editor | Inpainting nodes, SetLatentNoiseMask |
CONTROL_NET | A loaded ControlNet model | Load ControlNet Model | Apply ControlNet |
UPSCALE_MODEL | A loaded upscale model | Load Upscale Model | Upscale Image (using Model) |
CLIP_VISION, CLIP_VISION_OUTPUT | An image encoder, and one picture encoded by it | Load CLIP Vision, CLIP Vision Encode | Style, unCLIP and image to video nodes |
SIGMAS, SAMPLER, GUIDER, NOISE | The parts of an assembled sampler: the noise schedule, the algorithm, the guidance, the starting noise | BasicScheduler, KSamplerSelect, CFGGuider or BasicGuider, RandomNoise | SamplerCustom, SamplerCustomAdvanced |
VIDEO, AUDIO | A finished clip, a sound | Create Video, Load Video, Load Audio | Save Video, Get Video Components, audio nodes |
INT, FLOAT, STRING, BOOLEAN, COMBO | Field values: whole number, decimal, text, yes or no, a choice from a list | Primitive and math nodes | Any field of that type turned into an input |
Typical mistakes
- Trying to save a
LATENT: decode it first with VAE Decode. - Feeding a video sampler the empty latent of a picture: the type is the same
LATENT, so the link attaches, but a video needs a latent with a length. Use the empty latent or start node of the video family. - Plugging a
CONDITIONINGmeant fornegativeintopositive: the sampler then moves away from what you asked for. - Expecting a pack's type to fit a built-in input:
WANVIDEOMODELis notMODEL.
How the app helps
Select a link on the canvas and the Explain tab describes its type in plain words. Check names any link whose types do not match.