ColorizeDepthmap


Diffusion-based monocular depth estimation:  
https://github.com/prs-eth/Marigold  
  
- denoise_steps: steps per depth map, increase for accuracy in exchange of processing time
- n_repeat: amount of iterations to be ensembled into single depth map
- n_repeat_batch_size: how many of the n_repeats are processed as a batch,  
if you have the VRAM this can match the n_repeats for faster processing  
- model: Marigold or it's LCM version marigold-lcm-v1-0  
For the LCM model use around 4 steps and the LCMScheduler  
- scheduler: Different schedulers give bit different results  
- invert: marigold by default produces depth map where black is front,  
for controlnets etc. we want the opposite.  
- regularizer_strength, reduction_method, max_iter, tol (tolerance) are settings   
for the ensembling process, generally do not touch.  
- use_fp16: if true, use fp16, if false use fp32  
fp16 uses much less VRAM, but in some cases can lead to loss of quality.  


MarigoldDepthEstimation


Diffusion-based monocular depth estimation:  
https://github.com/prs-eth/Marigold  

This node is experimental version that includes optical flow  
for video consistency between frames.  

- denoise_steps: steps per depth map, increase for accuracy in exchange of processing time
- n_repeat: amount of iterations to be ensembled into single depth map
- n_repeat_batch_size: how many of the n_repeats are processed as a batch,  
if you have the VRAM this can match the n_repeats for faster processing  
- model: Marigold or it's LCM version marigold-lcm-v1-0  
For the LCM model use around 4 steps and the LCMScheduler  
- scheduler: Different schedulers give bit different results  
- invert: marigold by default produces depth map where black is front,  
for controlnets etc. we want the opposite.  
- regularizer_strength, reduction_method, max_iter, tol (tolerance) are settings   
for the ensembling process, generally do not touch.  


ComfyUI-Marigold

Related Workflows

Version

✅ Install successful

0.19s

22.11s

Nodes (5)