wanx-troopers

Color Matching

color-matching-choices

kj-color-matching.webp

GH:kijai/ComfyUI-VideoColorGrading

2.5D Tools

Wan Alpha

Kijai has adapted Wan Alpha “DoRA”: HF:Kijai/WanVideo_comfy:LoRAs/WanAlpha

Decoder needeed for WanAlpha: decoder.bin in the following locations (files have different hashes but same size..)

I did not know they originally had 1 file but split it for comfy to 2; ah it’s just the fgr (foreground?) and pha (alpha?) split into two files

Test workflow:

wan-alpha wan-alpha

Loops

so if anyone want to use the loop for nodes, do not disable comfyui cache like i do, wasted 30mn figuring those nodes need the cache

GH:yolain/ComfyUI-Easy-Use

Drozbay:

Loops are possible with the current execution flow but are still somewhat fragile and they don’t allow for starting/stopping partial executions. You can’t stop half way through a set of loops, change something for the next iteration, and continue. Indexing with lists is also not super reliable right now. Overall it’s often times easier and more stable to just lean into the practically infinite canvas and just make gigantic workflows. They are large but to me they are simpler to understand than having everything hidden in loops or layers of subgraphs.

Resolution Master

GH:Azornes/Comfyui-Resolution-Master

NAG

Hmm.. what is NAGGuider from NAG?..

Combating Video Contrast Drift

NormalizeVideoLatentStart

ComfyUI native now has NormalizeVideoLatentStart node which has been lifted out of Kandinsky-5 original implementation. The node apparently homogenizes contrast and color balance inside the video.

mean/std normalization applied when using I2V

kj-normalize-video-latent

WASWanExposureStabilizer

GH:WASasquatch/WAS_Extras contains among other useful nodes WASWanExposureStabilizer intended for a similar purpose

Pose Retargeter

GH:AIWarper/ComfyUI-WarperNodes

Video Blending From Fragments

kijai/ComfyUI-KJNodes contains Image Batch Extend With Overlap which can be used to merge together original video with its extension done using I2V or VACE mask extension techniques. Example of it being used in a LongCat wf: extend-with-overlap.

WanVideoBlender from GH:banodoco/steerable-motion is an alternative.

See also the next section on Trent Nodes

Trent Nodes

TrentHunter82/TrentNodes contains Cross Dissolve with Overlap node as well as WanVace Keyframe Builder and other nodes for examining videos, taking last N frames, creating latent masks and VACE keyframing. See also: Qwen Edit - VACE.

Hunyuan Video Foley

github.com/phazei/ComfyUI-HunyuanVideo-Foley

HF Space safetensors
ComfyUI-HunyuanVideo-Foley hunyuanvideo_foley_xl
ComfyUI-HunyuanVideo-Foley synchformer_state_dict_fp16
ComfyUI-HunyuanVideo-Foley vae_128d_48k_fp16

More Foley-s

Acestep XL SFT can generate stunning music.

Inside of Comfy you could Use Stable Audio or ACE… but tbh both are not that good

Hiding In Plain Sight

Ckinpdx

Ckinpdx a passionate AI artist has shared GH:ckinpdx/ComfyUI-WanKeyframeBuilder repository.

Ckinpdx Wan Keyframe Builder (Continuation)

which provides Wan Keyframe Builder (Continuation) node. This node was originally intended to prepare images and masks for VACE workflows. When SVI 2.0 was released the node was updated to facilitate workflows combining VACE keyframing, extensions and SVI references. The node has two distinct modes of operation: when images output is used and when svi_reference_only output is used. The modes are toggled by a boolean switch on the node.

Sample wf.

Ckinpdx Load Audtio And Split

ck-load-audio-and-split-1 ck-load-audio-and-split-2

Use this node to split audio between generation runs which produce various parts of the video with HuMo. Use Trim Audio Duration as shown to remove duplicate part of audio before re-assembling the video.

Other Repositories by Ckinpdx

Assemble/Disassemble

Assemble separate images into a sequence

image-batch-multi

Disasseble equence into separate images

get-image-or-mark-range-from-batch

Execution Timer

from GH:PGCRT/CRT-Nodes:

timer

Unilumos

UniLumos is an AI model for relighting a video. Workflows:

Voice Generators

MysteryShack:

Only 2 right now can cover all the languages people need… vibevoice and omnivoice. Sadly fish doesn’t support this.

momo1321:

Yes sadly fish audio not supporting many language beside English,I use omnivoice it can exactly copy the emotion,I use it in combination with index tts for the emotion control

Music/Song Generator

ACE Step/Acestep, alternative merges: HF:Aryanne/acestep-v15-test-merges/tree/main/xl

Misc

Frame Interpolation

Moved here.

Pose Detection

vitpose can do animals as well as humans.

dwpose

sd-pose

workordie:

SD pose is great it’s just slow in my experience

Q: sd-pose … taking a long time
[djbfilmz] A: your SD pose is probably not optimzied, also there is DW Pose (Tensor ver) which is really fast. I switch between .. I’m on a 5000s series card tho yolo_for_multi_sdpose_detection
sdpose
GH:judian17/ComfyUI-SDPose-OOD

Kijai:
!kj-sdpose

if you have a ton of VRAM you can use very high batch size to make it faster, but you can’t do that with cropping since cropping has to be frame by frame

Sapiens2

GH:facebookresearch/sapiens2

the cropping/bbox is only really necessary if your subject is super small on the frame, or if you want to detect multiple people

sapiens2 0.4B twice as fast as SDPose … they also have 0.8B and 5B models, and one 4k model

Q: How big is the vid, res wise?
A: 512x1024, the model does 768x1024 only anyway

sam3d-body [rainbow shapes???] is on another level

the 1b is a lot better than the 0.4, which totally failed this test

4k model is only 1b

Experimental implementation for ComfyUI: GH:kijai/ComfyUI-Sapiens2

sapiens2-nodes

Q: DWpose … deprecated now, considering we have Sapiens2? A: I’d keep it around. Sapiens looks really good tho

Sam3.1 And Sam3D

Preview of future nodes to come utilizing Sam3.1 for segmentation and Sam3D for mesh reconstruction: sam3d-wip; sam3d-object.

sam3d-object … it will be able to do whole scene too and then sam3d-body plugs into the scene

Q: it does do fingers tho?
A: sam3d does, but kimodo doesn’t; sam3d can even do ASL [American Sign Language], it’s really good; … did even add HAMER to refine the hands, but I’m not sure it’s … necessary anymore

Sam3 Native Model Loader node has been spotted in the wild.

Sam3 masking wf from djbfilmz: sam3-masking

This site is open source. Improve this page.