OpenArt
OpenArt
Workflows
Go to OpenArt main site
Upload workflow

8 Steps Style and Face Transfer with Unsampling and RF Inversion

3.7

3 reviews
46
11.4K
4.3K
9
Description

Table of Contents

Functionality
Demo

This workflow combines the power of RF Inversion with Alimama’s Turbo LoRA, enabling lightning-fast, 8-step image generation with unprecedented flexibility.

💡For a detailed walkthrough, watch the Youtube video tutorial:  https://youtu.be/H_G2AaLWN2o

B站主页：https://space.bilibili.com/3546611913329493

Functionality

  • Semantic Style Transfer: Transfer styles between reference images and generated images, including photorealistic, manga, illustration, and painting styles.
  • Pose Imitation: Recreate complex poses from reference images, going beyond what traditional tools like ControlNet OpenPose can achieve.
  • Natural Lighting Adjustment: Refine lighting and color in images for smooth, realistic effects without manual editing.
  • High-Quality Image Upscaling: Transform low-quality images into high-definition versions while retaining essential details.
  • Face Consistency across Images: Maintain the same facial features across multiple generated outputs for cohesive storytelling.

Demo



Node Diagram
Discussion
H
Hidalgo Serra5 months ago

Hello

As soon as i am in the sampling section and the sample custom advanced node get loaded i receive this error:

SamplerCustomAdvanced

expected scalar type Half but found BFloat16

FLUX.1 Turbo Alpha

D
Dominique B5 months ago

have you changed anything in the settings compared to the original workflow, it works perfectly for me

H
Hidalgo Serra5 months ago

nothing, i just put the models that are not directly loaded in the workflow manually, but keeping the same one, only thing i did is to input two images, one for the reference and one for the face, should these images have specific size?

The issue begin when the sampling area of the workflow start

(Edited)
H
Hidalgo Serra5 months ago

i managed to have the workflow working, but the images comes really far from the reference image, face is very close to the reference, but the images looks very cartoonish compared to the reference and the face, any way to improve that? i already amped the steps a bit up to 12 and add 2 extra loras, but result is very similar.

p.s.

the YouTube channel is either showing a different workflow or a dated one, because main models and some part of the workflow are totally different

👍1
D
Dominique B5 months ago

Yes, i can't have very good results too

W
Wei5 months ago

Thanks for your feedback! The settings for eta, end step, and your prompts can really shape the outcomes of your workflow. Have you considered using more realistic images? Adjusting your prompts to emphasize a realistic style could help you achieve the results you're looking for. Give it a try!

H
Hidalgo Serra5 months ago

thank you for your prompt respond, the image i have put are photos, but i attach here the result, so maybe you can give a bit more info of what you are actually changing to get a more similar result to the refererence and a more realistic too

https://imgur.com/a/isod2Em

y
yogeshnichal5 months ago

Hands not good even using flux hand v2

W
Wei5 months ago

You can add facedetailer to fix hand, that's not important

Author

W

Wei

20
176.2K
1.4K
407.6K

Reviews

M

Max Horowitz-Gelb

5 months ago

Ignore Hidalgo. This pipeline works great. Make sure you download all the correct models. The key is using PulID for face embedding. Make sure Pulid detects your face. Also you may need to mess with the Flux Reverse ODE sampler node. Set eta trend to constant. eta to 0.7 and mess around with the number of steps. The more steps the more your output will look like your input image. Also set basic scheduler steps to 28 for better results.

H

Hidalgo Serra

5 months ago

Ignore Max Horowitz-what ever, he have tested is based on nothing, i have tested the model on different machines, i already have have all the exact models that the template require, messing with the settings will make no difference, the results are far from the examples and I don't mean to be rude to the creator, he did it for free, but the result on the workflow are horrendous, deformed, cartoonish and overall far from the result shown

W

Wei

5 months ago

Please watch my video tutorial

D

Dominique B

5 months ago

working fine !

Versions (3)

  • - latest (5 months ago)

  • - v20241029-091018

  • - v20241028-084239

Primitive Nodes (32)

Anything Everywhere3 (1)

ApplyPulidFlux (1)

Fast Groups Bypasser (rgthree) (1)

FluxDeGuidance (2)

FluxForwardODESampler (1)

FluxReverseODESampler (1)

GetNode (10)

ImageConcatMulti (1)

InFluxModelSamplingPred (1)

Label (rgthree) (1)

OutFluxModelSamplingPred (1)

Power Lora Loader (rgthree) (1)

PulidFluxEvaClipLoader (1)

PulidFluxInsightFaceLoader (1)

PulidFluxModelLoader (1)

SetNode (7)

Custom Nodes (29)

Comfyroll Studio

  • - CR Text (2)

ComfyUI

  • - SamplerCustomAdvanced (2)

  • - BasicGuider (2)

  • - DisableNoise (2)

  • - BasicScheduler (2)

  • - FlipSigmas (1)

  • - CLIPTextEncode (2)

  • - VAEDecode (1)

  • - PreviewImage (4)

  • - VAEEncode (1)

  • - ImageScale (1)

  • - LoadImage (2)

  • - UNETLoader (1)

  • - DualCLIPLoader (1)

  • - VAELoader (1)

ComfyUI Easy Use

  • - easy imageSize (1)

ComfyUI Essentials

  • - SDXLEmptyLatentSizePicker+ (1)

WAS Node Suite

  • - Image Flip (1)

  • - Image Crop Face (1)

Checkpoints (0)

LoRAs (0)