⬇️ Description below Requirements ⬇️
Requirements/Resources:
Custom-Nodes
Models
qwen_image_layered_bf16.safetensors
in models/unet/Qwenqwen_2.5_vl_7b_fp8_scaled.safetensors
in models/clip/Qwenqwen_image_layered_vae.safetensors
in models/vae/Qwen
Description
This is a simple Qwen Image Layered workflow.
This model has only been out for a few hours so it's not heavily optimized yet, therefore i am only releasing this workflow since there are no other workflows currently released on civitai or elsewhere.
If you want to find out more about this model you can check out the official huggingface repository, but what it basically does is that it turns your image into multiple layers for further editing in software like photoshop or even comfy itself.
This way you can remove a character and/or objects from the background and edit or compose them one layer at a time.
This Workflow currently ONLY supports seperating the image into layers - this is not an All-in-One Solution with Qwen-Image-Edit integrated to also edit the selected layers.
This Workflow is only meant as a simple and easy tool to test out the new model and use it to seperate image layers and is relatively barebones in it's functionality.
Neither the workflow nor the model are really optimized for low vram, all i can say is that it works well on 24GB of VRAM and can seperate an image into 3 layers in ~200 seconds at a resolution of 1024px.
Thank you!
Thank you for using this workflow and if you end up creating something with it i'd be excited to see your results posted here on the model page!
If you have any questions or problems feel free to ask them in the comment section here on this model page - but if you're having issues please remember to attach your error log.
I will probably update the workflow once i have played around with it more myself since i'm planning to use this for Live2D animation generation.
No promises though!
Description
FAQ
Comments (6)
The subgraph loads with a missed connection at the end - I think should go from the sampler to the decoder. I can't get this to run anyway because of tensor mismatches. Not sure what's going on, but doesn't work for me.
Also doesn't work with the qwen layered vae either for some reason. I thought I had a corrupted vae and downloaded it again but still didn't work?
Hey @EricRollei21,
could you show me what error your console is showing?
Make sure ComfyUI is on the newest version since 0.3.51 added nodes that are used here.
Thanks for the offer to help. I kind of went nuts and made myself a whole set of custom nodes to run this model that uses the diffuser pipeline since it's got more options like using the embedded vision model for captioning and naming the output layers. I also made a node to save the layers in a .psd file stacked with the original image. That's useful because Qwen-image-layered wants only 1mp images and so my nodes scale the input down to the correct qwen size (divisible by 32 near 1mp) and then scale the output layers back to the original input size for saving in the stack. If you want to explore that I put them up on github here https://github.com/EricRollei/Qwen_Layers_Diffuser_Pipeline_Comfyui I didn't optimize the nodes for lower VRAM (with block swaps or CPU offload) so they might not work for everyone. But you maybe able to make some workflows with those nodes that do a bit more.
@EricRollei21 that looks hella interesting and awesome to get it straight into psd files for editing, including the original base.
I'll check it out once chrismtas holidays are over and i was able to play around with them and probably update the workflow with some more in-depth functionality then!
Thanks for your contribution to the open source community! ♥️
I thought of trying this but then saw the model size :(
Hey @shikharisfree532,
i sadly have no control over the model sizes :/
But you can find quantisized models of it on the internet, they'll usually have something like Q4, Q6 or Q8 in the name.
They will be smaller but might also provide lower quality.