🌀 ComfyUI Workflow: Character Pose & Expression Generator for LoRA Training
Ever wanted to train a LoRA but only had one good image of your character?
That’s exactly what this workflow is for.
It takes a single reference image and generates your character in tons of new poses, angles, and expressions — all while keeping them on-model. Perfect for quickly building a clean, varied dataset.
What it does:
✅ Creates dynamic body poses (standing, sitting, kneeling, jumping, flying, punching, kicking, etc.)
✅ Adds anime-style expressions (happy, sad, angry, smug, love-face with heart pupils, confused, yelling, crying, etc.)
✅ Includes hand + finger gestures (peace sign, pointing, fists, clasped hands, heart-shape, etc.)
✅ Keeps the style consistent with flat 2D tones + crisp line art
✅ Gives you ready-to-train outputs with minimal cleanup needed
Why use it?
LoRAs really shine when the dataset has variety — but collecting or editing images of your character is a pain. This workflow basically does the heavy lifting for you: drop in your character image, hit generate, and you’ll get a nice spread of poses and moods to train on.
I made it to help myself, but I figured it might help others who are building LoRAs too. If you try it out, let me know how it works for you — or share your own tweaks back!
Description
Added the nodes for QWEN 2509
FAQ
Comments (19)
Новый уровень. Мое восхищение!
I'm asking a rather silly question to test my understanding of what LoRA does—can I use an image of a character I like and, through LoRA, generate outputs with specific poses, expressions, and actions as you described?
This workflow do it without a LoRa, it uses the QWEN edit model and create new poses from an image. It is not perfect and usually people use this kind of thing to select the good generations and create a dataset, so they can train a LoRa that will be better than this.
@alcaitiff Thanks for your reply. I haven’t officially looked into Qwen yet, because new AI models are coming out so fast that sometimes I can’t keep up with them despite spending time to learn.
@FLOW0308 I feel your pain =]
was about to remind you that QWEN 2509 got out and i didn't wait for you to update, here are some more tries from my parts :).
this time it's not just great !
IT'S MARVELOUS ! (not perfect but the model is so much beter)
I'll be building sets of characters with that workflow soon.
here are suggestion for improvements that i plan to make for myself :
. save the text which is used to generate the picture as a txt file next to the character (search and replace "make this person" to the provided chara name)
these are prompts already valid as captions.
. optionnal WD14 and Florence/LMStudio tagging.
. the controlnet face refine option would be tough but i'll keep that in my mind. some character won't need that so it would be bypassed most of the time
I saw on your previews images, that the character always wear the same outfit. Whitch is bad to train a character lora.
Does your WF generate a complete dataset with different outfits in each image?
I trained a character Lora for my own use, of one of my 3DX futanari comics. I used 43 images, with closeup face, medium shot, cowboy shot, full body shot, naked, clothed, etc...all with front, back, side views. And of course, in each image my character was wearing a different outfit.
As my character was a futanari, i'm a well know 3DX futanari comics artist/author, (just search Pat3dx on Google, or check Affect3d, subscriberstar, etc...)this required more image for training, as it need images of the futanari clothed, naked with flaccid penis, and erect penis, in a lot of views and closeup.
For a normal character, I would use only 10 -15 images, of course.
Does your WF do this?
Short answer: yes.
Long answer. Sometimes you want the model to learn an outfit like batman, superman, spider-man, you need several images with a costume, but sometimes you want only the face and body. This workflow does both. Its all about the prompt. There are 5 boxes inside the WF with prompt examples. One is called "Without costume", this one will add clothes to the char and it will be more generic. But the prompts are just examples, you can do your own.
always the same outfit is a valid design choice. training face is never as hard as training bodies and clothes which this WF provide a solution for.
training face with this workflow without forcing the clothes can be done by generating multiple outfit in prompt as he mention.
i would also suggest to add some turnaround view+face closeup+scenery merging to combat the stiffness.
QWEN EDIT 2509 can do that without much sacrifice
@alcaitiff Thanks for the answer :).
Yes, of course, i didn't mention it, but unless you want to train a specific character, like Mario, superheroes, etc...that always wear the same outfit. I just took my own example, as i did a training myself.
Is it possible to use more than one input image of the same character so that multiple face angles can be taken into consideration when generating?
It is but you need an image with the two views or change a little bit the wf.
Easy stuff!
1. add an image loader ( press alt + click and drag the image loader)
2. load on image
3. connect the image loader on the and connect on the TextEncodeQwenEditPlus at the Conditioning box
4. If necessary adjust the prompt
5. ????
6. Profit!
I don't have res_2s and bong_tangent among the sampler and scheduler options. How can I install them?
Maybe a dumb question but are the suggested prompts in the wf suited for creating realistic characters as well?
Sure, just clean some parts like:
Make this person jump high into the air with knees bent upward, both arms raised overhead. Eyes wide with starry sparkle pupils, mouth open in a huge joyful smile showing teeth. Fingers spread wide in excitement, white background, (remove this -> clean stylized 2D tones).
@alcaitiff Awesome, thanks for the quick reply and tips, ill give it a go!
nice ! update 2511 is sweet !
maybe a version with openpose and some closeup shot would do nicely