Marigold works best close to 768p

• denoise_steps: diffusion step count, more can increase quality, less is faster
• n_repeat: how many times the prediction is done
• regulazer_strenth/method, max_iter, tol: These are settings for the ensembler, generally do not touch
• n_repeat_batch_size: batches the n_repeats for speed, INCREASES MEMORY USE -use_fp16: can lower quality a bit, disabling DOUBLES MEMORY USE
• scheduler: different schedulers have bit different outputs, experimental

Model: Models are autodownloaded from Hugginface to ComfyUI/models/diffusers -folder To use LCM model use just a single denoise_step and n_repeat 4

You can use this to adjust the depth map range