ComfyUI 中的 Empty Latent Image 节点

Q: Error: height and width must be divisible by 8

A size like 1000x700 was entered. Use multiples of 8, ideally 64 (1024, 960, 896, 832, 768).

Q: Out of memory with a batch

batch_size multiplies the memory the sampler needs. Lower batch_size and queue several runs instead.

Empty Latent Image creates the blank latent canvas (width, height, batch_size) that text-to-image sampling starts from; dimensions must be multiples of 8 and match the model family.

节点包:ComfyUI core

分类:潜空间

此页面尚未翻译,因此以英文显示。

Empty Latent Image 的作用

Text-to-image needs a starting point. Empty Latent Image produces a latent tensor filled with zeros at the requested size, and KSampler adds the noise. Latent space is 8 times smaller than pixels for SD models, so a 1024x1024 image is a 128x128 latent; the node takes pixel sizes and divides for you.

Pick sizes the model was trained on: 512x512 or 512x768 for SD 1.5, 1024x1024 / 832x1216 / 1216x832 for SDXL, Pony and Illustrious. Larger first-pass sizes cause duplicated subjects; use a hires pass for more pixels. For Flux, SD3 and video models use EmptySD3LatentImage or the family-specific latent node, which has a different channel count.

输入

名称	类型	含义
width	INT	Image width in pixels, multiple of 8 (64 recommended).
height	INT	Image height in pixels, multiple of 8.
batch_size	INT	How many images to make in one run with consecutive seeds.

输出

名称	类型	含义
LATENT	LATENT	The empty latent for KSampler latent_image.

跳过安装:云端预装的 ComfyUI 工作流

BitVector 在自家 GPU 上运行现成的 ComfyUI 工作流。无需安装,没有缺失节点,没有红框。用手机或笔记本打开,一分钟出图。

试用 BitVector

如何使用 Empty Latent Image

Add Empty Latent Image.
Set width and height for the model family (1024x1024 for SDXL).
Set batch_size 1 while testing, higher once the prompt works.
Connect LATENT to KSampler latent_image and keep denoise at 1.0.

设置与技巧

SDXL sweet spots: 1024x1024, 896x1152, 832x1216, 1152x896, 1216x832, 1344x768.
SD 1.5 above 768 on the long side tends to double heads and bodies; upscale afterwards instead.
batch_size 4 is a cheap way to explore seeds; VRAM scales with it.
Width and height that are not multiples of 8 are rounded down.