ausboss
b160e8bcbc
Wired straight into a Qwen Image 2.1 sampler, the node's latent made a picture twice as wide and tall as the size shown (1024 x 768 came out 2048 x 1536). The latent now carries how far it is downsampled, as ComfyUI's own empty-latent nodes do, so ComfyUI fits it to a model that packs pixels differently. Other models are unchanged.
24 lines
2.8 KiB
Markdown
24 lines
2.8 KiB
Markdown
# Latent Size 🆎
|
||
|
||
Choose the width and height of a canvas visually, then wire those integers or the empty latent into your workflow.
|
||
|
||
- **Shape:** the left button selects landscape or portrait. Ratio chips follow that choice; square retains it for the next selection. Clicking a chip uses its generated size near one megapixel. Enable **Chips keep current budget** in the gear to retain the current area.
|
||
- **Canvas:** drag a side to change one dimension; drag the corner to scale both. Shift frees the corner ratio, Alt removes snapping, and Escape cancels the drag. The bright readout shows dimensions and megapixels. Arrows nudge the focused canvas.
|
||
- **Size & budget:** scrub or type W, H, or MP. Typed dimensions are preserved exactly. The swap button exchanges dimensions. Size chips offer larger and smaller versions of the current ratio.
|
||
- **Gear:** set the **Snap step**, edit the **Ratio rail**, choose the **Latent family** (`latent`) and **Batch size** (`batch_size`), and hide the **Dot grid** or **MP cost curves**.
|
||
- **Node size:** drag the node's corner to make it wider or taller, so it lines up with the nodes around it. The controls stretch with it, and the canvas preview takes any extra height. It can be as narrow as 320 pixels.
|
||
|
||
## Outputs and links
|
||
|
||
Outputs are **width**, **height**, and **latent**, in that order. The width and height inputs have separate sockets above the panel. Linking either locks the controls that would change both dimensions; the other dimension remains editable. Linked values are resolved during execution; the preview shows the locally stored dimensions.
|
||
|
||
The **Latent family** is **16ch** (8× downsample), **4ch** (8×), or **128ch** (16×). Select the family required by your model. A model that packs pixels differently, such as Qwen Image 2.1, still gets the size you set: ComfyUI resizes the empty latent to fit it. This is an empty *image* latent; video models should take the width and height outputs and construct their own audio/video latent.
|
||
|
||
Typed sizes need not be divisible by the latent downsample factor. The integer outputs retain the typed size, while the latent spatial dimensions round down to complete cells, matching the core empty-latent nodes. Use aligned dimensions when wiring the latent directly. Large dimensions and batches allocate correspondingly large tensors.
|
||
|
||
## Limits
|
||
|
||
Dimensions range from 64 to 8192 pixels. Extreme custom ratios may not have an exact solution inside that range at the chosen snap; they use a bounded approximation. Fine custom ratios can likewise only approximate the ratio during a snapped drag. Longer custom ratio lists scroll within the shape area.
|
||
|
||
All execution inputs are ordinary widgets, so API runs do not depend on this panel. Orientation and snap preferences are saved with the node. This node has no model downloads, network requests, or additional pip dependencies.
|