Finally, a Solution for Consistent Character Posing!
Simply input any character you want and the poses you want them to perform, and that's it.
The examples I gave aren't cherry picked. I've tested it like 30 times and they all give me exactly what I asked for.
While waiting for WAN 2.1 to release an image-to-image model with ControlNet support, my workflow is currently the best alternative for achieving consistent character poses.
Better and much faster than Flux Kontext.
On a A4000 gpu with 16GB VRAM, it takes around 40 to 50 seconds to generate 3 images.
Link for the fastest wan2.1 model that I'm using in my workflow: https://huggingface.co/QuantStack/Wan2.1_T2V_14B_LightX2V_StepCfgDistill_VACE-GGUF/tree/main
Disclaimer1 : I'm using a very optimized Wan2.1 vace model, if you look at the generation setting, I'm using only 4 steps and 1 cfg. If you want to use your own Wan2.1 vace model, increase those setting with whatever works best for your model. But I highly recommend downloading the one I linked because it's super fast!
Wan 2.1 vae : https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/vae
Disclaimer2: if you're getting triton/torch/cuda erros, bypass the torchcompileModelwanVideoV2 Node, then "update all" in comfy manager, then restart.
The clip model I'm using : https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors
Poses I used: https://civarchive.com/models/22214/openposes-collection (and ofc use your own poses)
Now you can use just one reference image to create a LoRA for your character,
or why even bother with a LoRA when you have this workflow?
Description
FAQ
Comments (23)
It would be great if you post a simple video result (mp4), like the first lady image.
what is the size of the input image?
The one I used in my example? 512*1150. But it doesn't matter the size of the input, it can be 4000*4000, or 320*730. the resize node later on make it smaller following the value you have in the "image resize" in the "to configure" group. which in my worflow is set at 600 I think. Higher value will give you higher quality results, but much slow speed.
you know hot to train Wan 2.1 Text to image LoRA? like character LoRA and style LoRA?
thankyou for your hard work...
alot of additional info..
https://www.reddit.com/r/comfyui/comments/1m5hc43/2_days_ago_i_asked_for_a_consistent_character/
ackend='inductor' raised: RuntimeError: Cannot find a working triton installation. More information on installing Triton can be found at https://github.com/openai/triton Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information You can suppress this exception and fall back to eager by setting: import torch._dynamo torch._dynamo.config.suppress_errors = True
what happened
Hi,
I've added the solution in the model descriptiotn because many peopel have had the same issue, if you're getting triton/torch/cuda erros, bypass the torchcompileModelwanVideoV2 Node, then "update all" in comfy manager, then restart.
face are super distorted is there anyway to improve it, same model/vae/clip/settings, increasing steps will cause extra items being added
connecting the ouput images to a facedetailer will fix the issue. It's very simple to do.
I was actually more interested in the reference images than the workflow—they look really good! Would you be able to share the original images with metadata, if possible?
There was nothing special about the workflow, just the default text to image with no lora.
I used this illustrious model: https://civitai.com/models/835655/illustrious-xl-personal-merge-noob-v-pred05-test-merge-updated
And the text prompt started with " cartoon style, dark outline, full body, standing" then a description of the woman .
perfect workflow ever
just spits out the reference images for the controlnet some reason lol
Hey, did you ever find a solution to that problem? im having the same issue despite trying various fix i still get the reference image as single output result.
nouvelordreops524 Never found solution, unfortunately
I had the same problem with an other downloaded workflow, I think this happens when you have the wrong checkpoint/controlnet/clip or any other bad model input. If you load the correct models in all the input nodes it should work just fine
Thank you! Can you please add face detailer also.
This is such a great tool. Do you think it's possible to get even better results with Wan 2.2? Or is 2.2 incompatible?
Can you make the same but with emotions please?
I load an image of a single person and the result I get is 3 people and none of them are posed correctly. I have no idea what to change. I made sure that the pose images are all exactly the same size/ratio.
Hello, I only get poses, what am I doing wrong?