This workflow enables you to generate transparent images with ANY checkpoint architecture that doesn't currently have a trained Layer Diffusion model for it. It doesn't magically make the existing models work with other architectures, but uses some clever tricks to go around these limitations. Contrary to my other work, this workflow is as basic as it gets so that you may incorporate its ideas into your own . Please read about the specifics below. Dependencies <a target="_blank" rel="ugc" href="https://github.com/huchenlei/ComfyUI-layerdiffuse">https://github.com/huchenlei/ComfyUI-layerdiffuse</a> <a target="_blank" rel="ugc" href="https://github.com/Fannovel16/comfyui_controlnet_aux">https://github.com/Fannovel16/comfyui_controlnet_aux</a> How it's done First we generate an image with a supported model. At the time of this post those would be SD1.5 and SDXL. Pony works too, but Illustrious/NoobAI do not. Here's an image generated with a Pony checkpoint and the SDXL Layer Diffusion model applied but before decoding Layer Diffusion (so no transparency yet): <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a156e2c6-8934-464f-be9c-a417a47c59fe/width=525/a156e2c6-8934-464f-be9c-a417a47c59fe.jpeg" />After we decode we get a transparent image. CivitAI doesn't really support that but the black here is actually the transparency. <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e170eb50-752d-4bd0-91eb-57599589a21d/width=525/e170eb50-752d-4bd0-91eb-57599589a21d.jpeg" />This is still the Pony image, so something you can do without any fancy workflows. We want to generate on a checkpoint that doesn't have a trained Layer Diffusion model. So instead of saving this, we take the mask from the Layer Diffuse Decode node: <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a9efb801-a8c4-44c8-aa09-6b3970196dc7/width=525/a9efb801-a8c4-44c8-aa09-6b3970196dc7.jpeg" />Now we take the very first image we generated and use ControlNet preprocessors to create depth and HED (soft edge) images based on that Pony image before LD decoding (lineart can also be used instead of soft edge): <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a6f0688d-c57f-4804-9933-4a1f5ecfbf0f/width=525/a6f0688d-c57f-4804-9933-4a1f5ecfbf0f.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b88d4103-0969-48c6-9234-569d1d6a18e1/width=525/b88d4103-0969-48c6-9234-569d1d6a18e1.jpeg" />Applying both ControlNets we generate an image with the desired checkpoint that we don't have a Layer Diffusion model for, here an illustrious checkpoint: <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6661cb4a-f404-45e8-9988-4ccb335c1444/width=525/6661cb4a-f404-45e8-9988-4ccb335c1444.jpeg" />We need to set the ControlNet strengths sufficiently high so that the character stays withing the bounds of the mask from earlier. Now the last step is to crop out the character using that mask: <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d5dc7573-2ce5-43c2-b9f9-e470005c6887/width=525/d5dc7573-2ce5-43c2-b9f9-e470005c6887.jpeg" />Viola! A transparent background image with Illustrious but without an Illustrious Layer Diffusion model. Checkpoints used: <a target="_blank" rel="ugc" href="https://civarchive.com/models/317902?modelVersionId=593760">T-Ponynai3</a> <a target="_blank" rel="ugc" href="https://civarchive.com/models/1052681?modelVersionId=1208830">Aungir</a> LORA If you want character loras with this, you might need to have one for both architectures. Same may be the case with pose models, but more se on the one that's earlier in the workflow. Upscaling To add upscaling, you'll want to do it before cropping the image. From my tests, you don't need to upscale the mask, it will get scaled automatically (in fact when I tried scaling it, the results were worse). I tested with tiled upscaling and it worked very well.

Files with SHA256: 6bef79f391142b2408db1504e3840fef5c8dab036acfa0d523a5cf47fc14ffe3