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://civitai.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?