ComfyUI Pipelines

PipelineDesign · Build

Overview

Local AI video workflows, built in ComfyUI and run on my own hardware - custom segmentation, inpainting, and video generation assembled for production control.

Built the graphs stage by stage - Florence 2, SAM2, ProPainter, Wan 2.1 VACE - so a removal that can’t be reshot still lands inside a real post schedule.

This is the chain behind The Travel Writer - the Fiat Ritmo removal on Heidi Ewing’s second feature, built here and delivered into the cut on a five-month post schedule.

TypePipeline

PlatformComfyUI · Local

RoleDesign · Build

RITMO - the plate mid-pipeline: the car detected and matted

Toolchain

Florence 2Open-vocabulary grounding - one text prompt (“car”) instead of a training set; its boxes seed the tracker

SAM2Segmentation with memory - one box prompt and the matte propagates the whole clip, not per-frame guesses

ProPainterFlow-guided video inpainting - wants its masks dilated; a tight matte ghosts at the edges

Wan 2.1 VACEGeneration under control signals - the masks decide where it invents and where the plate stays

ComfyUI · Pipeline - object removal

Load Video (Upload)

IMAGE

videoritmo_0370.mov

frame_load_cap121

RITMO - the plate as shot, the parked car behind the subjects

Florence2Run

imageimage

text_inputdata

taskcaption_to_phrase_grounding

text_inputcar

RITMO - Florence 2 grounding: the car detected in frame

car 0.97

Sam2Segmentation

imageMASK

bboxes

keep_model_loadedtrue

RITMO - SAM2 matte: the car held through the move, subjects cut out

ProPainter Inpainting

imageIMAGE

mask

resize_ratio1.00

flow_mask_dilates8

WanVaceToVideo

control_videoLATENT

control_masks

length81

steps20