Details
Recommended Prompt:create turnaround sheet of this exact character, 5 full-body poses on pure white background: front view, 3/4 left, left profile, back view, right profile, 3/4 right — evenly spaced in a clean horizontal row, consistent lighting and style, no background noise, no shadows, same proportions and detailing across all views
*feel free to tweak the prompt, but make sure you have "create turnaround sheet of this character" in your updated prompt.
Notes
It won't work well on real human photos
I recommend inputting a full body shot rather than a headshot of the character as the input image
It's not good at creating facial expression sheet, but planning to train one later on for that purpose
Without the lora, my example prompt itself also generates turnaround sheets, but with LoRA it's way better
Trained with Ostris AI Toolkit
Description
FAQ
Comments (31)
Please share your workflow for working with this LoRA in ComfyUI.
It's a standart LoRA implementation. Add the Lora after your checkpoint.
I put two snapshots of the workflow here, one using Nunchaku and one for standard workflow: https://imgur.com/a/iBjcozx
@reverentelusarca Yes, but how do you get a finished canvas from different angles from one vertical image?
@Kappa_Neuro The comfy workflows are present in the example images. Drag and drop them in comfyui.
@Kappa_Neuro this is a flux kontext lora so it trained on pair of images. one image is single angle other image is multiple angles of the same image.
I got higher image reference results from increasing Steps a bit, and lowering lora strength to 0.2
@aih220 prompt itself, without the lora, already works okay, lora just guides the generation, so please feel free to experiment and share more results!
Is there a way to make it work without nunchaku? i'm having issues installing those nodes properly.
@EarthwormBob Look at the first comment by OP in this thread. Nunchaku isn't a requirement for this Lora, you can use it normally.
@nymical yes, you can use with any Kontext model workflow, just put a Lora Loader node as usual between Checkpoint loader and sampler
OK, all you must do if drag the ComfyUI-numchaku workflow image into ComfyUI and edit it a bit:
a) Don't download the ComfyUInumchaku workflow.
b) Add the Load Diffusion Model node and change the entry to be the flux1-dev-kontext_fp8_scaled.safetensors
c) Add the Load LoRA node and change the entry to be the kontext-turnaround-sheet-v1.safetensors
d) Route from the MODEL node on b) to the model node on c)
e) route the CLIP node on "DUALCLIPLoader to the clip node on inbound c)
f) route to outboud clip node on c) to the inbound clip node on "CLIP Text Encode"
g) route the outbound MODEL on c) to the inbound model on the "KSampler"
h) Now delete the 3 unneeded numchaku nodes & save your workflow
i) Load your image(s) and run the generation & you're done. It's very slow for me though.
Kappa_Neuro connecting to an empty latent?
What is the recommended resolution? The example images have 1344x768 pixels, which seems not to be a recommended flux kontext resolution.
I assume the input image of the character should be placed at the left, isn't it?
There's only one input image. So, there's no left or right.
If you're concerned about the image dimensions, they can be set manually, they don't have to be set by your input image.
@nymical this is not what he asked. He is confused where to place the input image, not between right or left. He also asked the dimensions you recommended, not if they can be set manually. I think this is a translation error here, you might be better off using a AI translator.
https://imgur.com/a/iBjcozx please check these example snapshots. i recommend any landscape aspect ratio for fitting in the scene.
@wilderop69 Yes, english isn't my first language so I might have misunderstood what was asked.
I'm not OP though, so resolution wasn't recommended by me, I was just trying to help until OP replied. Many workflows (including this one) have 2 input images which are then stitched together, and resolution is set by that stitched image. So, I just thought it was about that. Again, I might've misunderstood.
Perfect! I hope you do more like these in the future! :D Seems really useful!
Thank you! If you have any recommendations please feel free to let me know
For me it doesn't output consistent characters. Please share all the details. Res of the input, res of the ouput, guidance, str, sampler, etc
you can drag and drop the example images into the comfy ui, also here is a snapshot of the workflows im using: https://imgur.com/a/iBjcozx
strength:1
res of the input: doesnt matter because im resizing into the nearest SDXL resolution with a node. tried with vertical, square and landscape, all of them are working. make sure your output latent image size is landscape such as 1344x768
@reverentelusarca Thanks! Will try ^^
it works great, but should it be called Flux Kontext "Nunchaku" Character Turnaround Sheet LoRA..
you need this "svdq-int4_r32-flux.1-kontext-dev.safetensors" from the the mit-han-lab/nunchaku-flux.1-kontext-dev page on Hugging Face
you can use the lora with regular flux kontext workflow, without needing of nunchaku. there is no such thing as "nunchaku lora", it's just an acceleration method and it's optional
Kontext responds well when you have a preface like this about your character. I make pictures from my Cyberpunk 2077 character that I made a contact sheet from. I use this all the time, no lora needed.
"The woman named V, as defined in the reference contact sheet, is a lifelike, ultra-sharp, photorealistic character with long wavey black hair styled in a high ponytail with long parted bangs and a voluptuous, fit physique. Her exact facial features, hairstyle, and body structure must be preserved in all scenes, ensuring consistent depiction as a vibrant, confident, real-life person transformed from a computer-generated model into an authentic human presence with natural skin, realistic proportions, and lifelike textures. Any narrative prompt that follows should maintain these identity markers while placing V in varied settings, poses, or styles as specified, emphasizing exquisite details and intricate textures."
Hello, thank you for the good job. Just wanted to point that your prompt asks for 5 poses but you describe 6 poses on the further instructions. Not sure if thats important
I cannot for the life of me get nunchaku to install, through manager, pip install, etc.
same
use the nunchaku installer node to install nunchaku backend first. you can get that from searching custom nodes in manager. then it should install fine
куда её класть в какую папку???
How do I know what workflow (which nodes are connected to each other) to use for this?