Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

Consistent Character Creator 3.5 IS HERE! (Patreon)

Published:
2025-09-29 10:11:17
Edited:
2025-09-29 14:57:18
Imported:
2026-06

Downloads

Content

Recently, a new version of the Qwen Image Edit model was released, and it's better in every single way. It also allows for easier installation since we no longer need to rely on extra LoRAs for additional functionality. New features include:

Better consistency

Pose transfer from any image

Improved Virtual Try-On

Advanced upscaling & dataset creation

Speaking about the new Upscaling Setup...

I played around with this extensively and tested pretty much every tool that helps keep characters consistent during the upscaling process. The best one was still PuLID, but the problem is that it's an absolute pain to install and regularly breaks with new ComfyUI updates.

So I tested something else. I challenged myself to rely on native ComfyUI tools as much as possible. One such tool is USO (Unified Style and Subject-Driven Generation via Disentangled Reward Learning) by ByteDance. It was released natively for ComfyUI, but the problem is that it's not too good... on its own. But it turns out it works wonders for upscaling when you only want to change like 20 percent of your image data.

I'm still testing this section, so let me know how it works for you!

Using the Workflow

The workflow will automatically generate a set of optimized training images for any LoRA. But not only that—it will also automatically generate the entire upscaled dataset including detailed captions for you. Just drag and drop it into your favorite LoRA trainer like AI Toolkit or Flux Gym and start training.

Important: Run this workflow in two steps. First generate the dataset images, THEN activate the B-Dataset Creation group.

Setup Instructions:

Install custom nodes via the manager

Download the models (you can find all the links in the notes within the workflow)

Workflow Steps:

Import your character

(Optional) Add a prompt for your character—one simple sentence describing style and clothes

Name your character. This will be the output folder name and caption trigger word

Run the workflow

Remember: You can always tweak things and change prompts! Also, the workflow is modular—you can duplicate entire groups if you want more Virtual Try-On options, for example.

Speaking of the Virtual Try On section: For it and the Pose Transfer group you will need to import Reference images to work!

Activate the B - Dataset Creation group and let it run! You'll find your complete dataset with captions in comfyUI/output/YourCharactersName

Files

Comments

Peter Anderlund

Looks interesting! Will check it out! Thank you! 🙏

Mrs.LoRa

Ich freu mich schon den Workflow auszuprobieren 🤩...

Spooknik

Just got 3.2 running yesterday, what a nice shadow drop this is :) Quick question, shouldn't all of the seeds be set to the same int?

Sarya Catmander

the model linked points to old qwen edit instead of 2509, don't know if it is intentional or it was made to be used with 2509 or old one

Mrs.LoRa

can you also do one workflow with the full model please? or give some advise how to embed the full modell in the workflow? appreciated

Matthew Gale

New subscriber here - loving the content! This works great but I'm having difficulties with creation of a digital avatar of myself. The outputs are close... but the character does not look quite like me. I'm wondering if the workflow could easily be modified to have multiple reference images or use additional tech (pulid, instantid etc.) to extract identity features to ensure the subject matches? I could also see value in a workflow that takes multiple pictures of someone and outputs a high quality starter image that could be used as the input for the input for this workflow. I'm experimenting with a few things but would love to get your thoughts? Either way, highly appreciate your content!

JS56

A problem I keep having with all versions (including 3.5) is that if, after generating all the images, and when you move on to the upscale part (etc), if something happens to interrupt the process (ComfyUI crashing for example) you can't pick up frowhere you left off. The whole thing has to start from scratch. You can't , for example, bypass all the first section so the second (Upscale + Florence, etc) just works on the alreadpreviously generated images from before the process was interrupted.... At least, I haven't been able to do this. The thought of going through all the images generation stage AGAIN is quite off-putting. Don't get me wrong, your overall program is great and I was using version 3 a lot, albeit with the aforementioned problem happening now and again. I just tested your 3.5 version and found it to be superior in creating character likeness, particularly the close-ups though the distant shots still need some work. The Florence section and the Upscale section seem to be much much slower than version 3. The program actually crashed in the middle of (a very long) upscaling session, and now I have to run the whole thing again... In short, better likenesses on close shorts, not so much on longer ones, and horrendous lengths of time for Florence and upscaling.

Spooknik

This is working great for me. One small pro-tip. The TextEncodeQwenImageEditPlus node allows for up to 3 images. As long as they don't really conflict you can do some cool stuff. First image is a cropped face, second image is arm tattoos. It keeps both images consistent across the whole workflow. This works better than prompting "arm tattoos" because each one was a bit different even with the same seed.

11111

Great work! Thank you!

11111

Did you try new wan2.2 animate? What do you think is it possible to run on 5090? I tried raw launch and stuck on memory limit, full 14b takes about 60gb vram, even Ada 6000 won't handle it..

Göran Burlin

when I try the dataset creation I ran into an error VHS_LoadImagesPath directory is not valid

Spooknik

You need to add your own input image it seems like.

Göran Burlin

I have finishen the first part of the program and all of the images are in the new folder made by the workflow. So there is a bunch of images but they will not load into the other part of the workflow

Bill

Same here.

Lorenzo von Matterhorn

Check your Witelist-dirs.list in ComfyUI\user\default\was-node-suite. I had to change the path to use slashes instead of backslashes. Then it works for me.

Felipe Gonzalez

why every time this happens VHS_LoadImagesPath directory is not valid: ComfyUI\output\Test_01. thanks

Gurpreet Singh

Did you get this resolved? what's the solution

Gurpreet Singh

Did you get this resolved? I am getting the same error

Gurpreet Singh

Did you just change one slash after C: i.e. the drive or all backslashes to Slash

Beri

for me this is the worst version, bad likeness, turning characters into Asians or plastic dolls, refusing to place character into a requested pose, half of the nodes are not rendering correctly

Felipe Gonzalez

if I run just one area at the beginning and them the rest at once it works. but the last part, the one bellow still I can't

quynh

I think you can hide all the image generation (part A) by Ctrl-B except for the 1st input, the description, Lora name on top left. This will save you time

Lorenzo von Matterhorn

Try specifying the full path, but without the individual folders. For example: C:/Users/XXX/Documents/ComfyUI/ and C:/Users/XXX/Documents/ComfyUI

Lorenzo von Matterhorn

Above the "Preview Image" window, there's a field labeled "Name." Expand it and enter your full path (e.g., C:\Users\xxx\Documents\ComfyUI\output\Training Models). Then the query works for me. You might also want to click "Update Inputs" under "Join String Multi."

Göran Burlin

Thank you! That solved my problem! How did you find out that solution?

Sarya Catmander

use only output, if on Linux remember to change / or \ accordingly

Sean jones

any ideea what this is? popped up on the k sampling(after a long while) on the Dataset creeation, comfui crashed after trying again prompting me to start again? CUDA error: unspecified launch failure CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Badbutton

Really happy I subscribed. I downloaded this and grabbed all the models needed and thought I'd slap in an image to test it out. Took a minute or two to produce anything and then all the following images were spot on. Fantastic workflow, worked right out of the box!! This is going to be a goto tool for me when creating a character chart.

Jeremy Bedingfield

Ok So Yes me again... Just totally re-installed Comfy, and installed all aspects of this again and from fresh. And wow does this work all of a sudden. The new 3.6 adv is amazing and worked first time and exactly as it should so ignore my previous comments amazing, thank you.

Nathaniel Clutter

Hi Mick, I uploaded a photorealistic reference image to character consistency 3.5 and was hoping to make the character into anime style. I took crystalwizard's advice on discord and put "anime style" as the first 2 words, but the images come out very oddly. Some are photorealistic, some are anime, most of them are blurry, and the facial expression images look very warped. As far as I know, I've followed your directions. Does anyone have any insight was to whether I just need to change my prompt, generate photorealistic photos before making them anime, or do something else? (I'm trying to figure out how to add screenshots on here or to you on discord)

Stan

look at the path and add the drive letter if it is not in the string. I have seen that a few times on the different releases.