⬡ KSampler¶
Central hub node: receives models, settings, prompts, and images — samples and produces the generation pipeline.
Inputs¶
| Name | Type | Required | Description |
|---|---|---|---|
model_bundle |
UME_BUNDLE |
✅ | Model + CLIP + VAE from a Loader node |
settings |
UME_SETTINGS |
✅ | Parameters from Generation Settings |
positive |
POSITIVE |
✅ | Positive prompt text |
negative |
NEGATIVE |
❌ | Negative prompt text |
loras |
UME_LORA_STACK |
❌ | LoRA stack from LoRA Block nodes |
image |
UME_IMAGE |
❌ | Image bundle for img2img/inpaint/outpaint |
Outputs¶
| Name | Type | Description |
|---|---|---|
gen_pipe |
UME_PIPELINE |
Complete pipeline with generated image, ready for post-processing or saving |
What Happens Inside¶
The KSampler orchestrates the full generation pipeline:
- Validates the model bundle (model, clip, vae must all be present)
- Applies LoRAs from the stack (if connected)
- Encodes prompts via CLIP (with caching for repeated prompts)
- Applies ControlNets from image bundle (if present)
- Prepares latent — empty for txt2img, or VAE-encoded for img2img/inpaint
- Samples using the configured sampler + scheduler + steps
- Decodes the latent to pixel space via VAE
- Packs everything into a
UME_PIPELINEfor downstream nodes
Optimization
The KSampler caches prompt encodings and ControlNet models. If you change only the seed, re-encoding is skipped for faster iteration.