One photo in. Full character sheet out.
Drop in one photo of someone and you get back a clean 3x2 sheet on grey: her face up close, plus five body shots. Same person, same outfit, every box.
Two modes: angles (front, back, side, two three-quarter turns) or poses (side-on, sitting on a cube, foot up, looking back, sitting on the floor).
How to use it
- Grab the workflow zip from the files on this page and drag it into ComfyUI.
- Load your photo. Any size. The workflow pads it for you (you'll need the free KJNodes pack).
- Paste one of the prompts below, exactly as it is.
- Hit run. A box looks off? New seed.
The stuff that matters
- Use a standing photo, head to at least mid-thigh. Sitting or close-up shots break it.
- Base Qwen Image 2.1 only. Turbo wrecks the grid.
- LoRA at 1.0, 25 steps, CFG 1. Go lower on strength and you get empty boxes.
Angle sheet (start here, it's the strongest)
turn this photo into a character reference sheet: a 3x2 grid on a plain grey background with a close-up portrait of her face, then her full outfit from the front, back, side, three-quarter front and three-quarter back, cropped at the neckPose sheet
turn this photo into a character pose sheet: a 3x2 grid on a plain grey background with a close-up portrait of her face, then her posing standing side-on with her back arched, sitting on a grey cube, with one foot up on the cube, looking back over her shoulder, and sitting on the floor leaning back, cropped at the neckWant sharper eyes? The second workflow in the zip upscales the sheet and copies her eyes from the original photo. Two more prompts and the full guide are in the zip too.
Trained with Fizgig on 77 before/after pairs, epoch 9. Follows the Qwen Research License (non-commercial).
Description
First release. Angle sheets are the strong mode. The workflow pads your photo for you.
FAQ
Comments (16)
This was fun to make, you need the workflow
I trained on FizGig, and our outputs were 16:9, so without the workflow, you will MOST LIKELY have a broken output
Will improve in future potentially.
Just a fun forty-minute project. Enjoy!
Before I try it later, how does it perform with 3D render and 2D illustrations & drawings?
Not in the training data, but I just tested it, they mostly work well. I think in future, I will cut out some of the "poisoned" data from ChatGPT. I let some slide, as this was just a quick run.
Update: Ran the tests and updating the gallery. Seems to work, i hope it works well for you, thanks @cliang96844
Great LoRA, but it's missing a close-up profile shot.
Noted! Thanks for the feedback, i can definitely train a new version soon
Works fantastic, tried characters from photo, AI characters and 3D rendered characters, it does a a great job, the only situation it gets confused when I load in large BBW female model 3D characters with massive butts, big thighs, large bellies, it does not know what to do so it generates random shapes and figurines that does not look like the actual mode. Overall this is a fantastic Lora and I recommend it to everyone, easy to use.
Thats extremely kind!
I must admit, i think i will retrain it with less grid images, maybe 3 or 4 maximum, and then trim out any of the Ai looking face images to increase quality of the dataset.
I will search for some BBW characters to try to satisfy
Will be implementing face-crops, expression sheets, etc.
Again, this is just a fun project, so i appreciate your kind words!
@ladaagency Considering that this was just a fun project, it turned out very well. Take a look at my deviantart gallery, user grimmspring and see if you find something useful out of bbw and vouluptuous shapes for your future projects, or just google up grimmspring and you can find some voluptuous and bbw shapes that I was talking about.
Thank you so much for sharing this amazing workflow for free! I did notice that it can be a bit unstable when rendering anatomy/NSFW details, so I'm trying to find a matching LoRA to fix it. But as you know, there are barely any LoRAs for Q21 in that department yet. Correct me if I'm wrong though! Would love to hear your thoughts.
I must admit its training data was from GPT edits, I don’t have any NSFW in there.
I may add some in future. I probably will.
That being said, there’s lots of NSFW lora, just combine them with this Lora
@ladaagency Thanks for the reply, and huge thanks for sharing this absolute beast of a workflow!
I'm wondering... how did you make the nice example promotional images with the red/black background, text and stuff? What software/tools did you use? Photoshop? Or did you use ai for the design?
Hi I had Claude create some designs, once I got one that I liked I just had it save the styling
@ladaagency Thank you for elaborating. In what format did you have it generate design (and save the styling)? It really looks good and professional.
@Remiel Ive not a clue, its saved as an MD somewhere inside my "Claude Brain". Im technical but lazy, i'll just have it find it when i need it lol