Using the power of Qwen Edit 2511, repose a reference character to a new pose to your heart's content!
All the relevant model DL info can be found within. Minor notes scattered about.
Its very easy to use. Put in a pose image. Put in your character sheet, preferably made with my Lazy Character Sheet workflow. Then just hit run. The loras and 2511 are extremely flexible and it seems a great variety of poses can be executed.
Note that the character in the pose image determines the style, body type and even expression of the final output. If you want a specific style, body type or expression, generate that first.
Version 4.0
Total overhaul
Model is now Qwen Edit 2511 GGUF version
Using the wonderful pose extraction and repose loras from tori29umai (details here)
Simplified everything
Version 3.0
The workflow now uses the Qwen Edit 2509 All-in-One model by Phr00t, found here.
This simplifies the workflow
Also, there's now both a SFW and NSFW model option
Notes updated accordingly
Version 2.0
Now the workflow has been rejigged and no longer uses the repose lora
The model has been updated to the Qwen Edit 2509 model
The workflow now uses the new model's Control Net features using DW Pose
New Image concat outputs an info image with inputs 1, 2 and the output.
Version 1.0
old and can be ignored
Notes:
If you are getting problems like this, try increasing the edge buffer or re-generating the first portion. The edge buffer is by default 5% and the seed is fixed by default.
Cheese and have a good fun genning!
Brie Wensleydale ~
Description
This is the 4.0 version adjusted for the AIO model which allows for NSFW access.
FAQ
Comments (16)
It was generated several times. Then the error began to appear - VAEDecode free_upper_bound + pytorch_used_bytes[device] <= device_total INTERNAL ASSERT FAILED at "C:\\actions-runner\\_work\\pytorch\\pytorch\\pytorch\\c10\\cuda\\CUDAMallocAsyncAllocator.cpp":563, please report a bug to PyTorch. All updates are worthwhile. 4090 24Gb + 64RAM. I have restarted ComfyUI and restarted the computer, but nothing helps. Could the problem be that the scan resolution of the four body positions is too high, which is loaded into the "character sheet" node?
I have no idea mate, I have not seen that error before. However, we do have nearly identical systems and I have similar errors before.
I did a stealth update and replaced the GGUF loader node with the UNET Loader (GGUF), that maaaaybe might fix your issue.
Also, another thing you can try is the another GGUF model. Q6 is faster and BF16 is even better quality. I currently run the BF16 when I want quality. I've personally had issues later on with the Q8 with that loader and I haven't tested it again since I switched to the new loader.
Within the tori29umai/QwenImageEdit2509_LoRA repository, I was unable to locate a LoRA model file with the name “repose”.
Friend, the links to both are in the notes, and the name should be loaded in the lora node.
One is named image2body and the other is named charaBG.
image2body
https://huggingface.co/tori29umai/QwenImageEdit2509_LoRA/blob/main/QIE_image2body_V2_dim4_1e-3-000078.safetensors
charaBG
https://huggingface.co/tori29umai/QwenImageEdit2509_LoRA/blob/main/QIE_charaBG_V8_dim4_1e-3-000048.safetensors
Works really well thank you! For the AIO version, where do I place the background?
This workflow is just for the pose.
I have an older Qwen Edit Character Fusion workflow that has the background part, but I haven't updated that yet. I just don't have the time right now...
@slipperygem Gotcha! It's fairly easy to add the background in after the fact anyway. Thanks so much for this pose workflow, gamechanging for my projects!!
@owhothot360635 Yeah, its a plain white BG, so its easy enough to remove.
For some reason, the loras are less stable when used with a BG on 2511, so its a bit more of a puzzle. My old Character Fusion workflow, while it worked, was too complicated and large. If I try to do the BG again, I want to find an easier way.
For a pose like this, you might want to reduce the edge buffer to 5 or lower. Then run it a few times and see if it sticks.
is there a discord where we can progress faster by posting issue etc? rn im testing it and wonder if you already have a small discord where we can pull all issue with it etc? if no do you allow me to make one and post it HERE?
I think enough people have asked me to make a Discord that I will do so once I figure out how. I'll post it in early Feb. Cheers.
@slipperygem i struggle making it work check the screenshot if you could help that's why i asked for a chat plateform.. https://mega.nz/file/Hm4S0KLJ#mN_Sw5b5Ek1eu5Nd4bs63XBxo3wxeGBNLZFXtEMSB4E
@carlitajohn89931 Yeah, I DL'd the file, I'll try to help int he DMs.
Thanks for the WF and the tori29umai's links... i guess than other than the 2 loras you picked, all the other loras on that page are redundant on qie2511?
Can't get it to run. Sam3 gets hung on install. Poking around to try and figure out why, but I just can't get to to fully install. Doesn't matter if i do it manually , or from within comfyui.
EDIT FROM THE FUTURE... THAT'S NOT THE PAST......
I managed to get it working. I found an issue on the sam3 issue tracker where person mentioned a fork that worked. You can find it by googling "wonderstone ComfyUI-SAM3 github" should be one of the first few results. I manually installed that node-pack and that fixed the issue.
The issue on the git said something about support for .pt files having been dropped at some point. Which matches up with the error i was having, which is the (Down) SAM3 Model loader node wasn't finding the node. As well running the WF with whatever comfy defaults to breaks the WF.
Anyway hop this helps someone. o7