But... why?
I created this workflow because I thought I had pushed the limits of what Hunyuan 3d-2 could do, and someone stepped in and proved me completely wrong. Because of this, I fell through the rabbit hole in discovering what hy3d-2 was even capable of. I believe that even their demos did not understand that their quality could be significantly better.
Base workflow
The base of this workflow belongs to Kijai. Please visit his github for instructions on how to get started.
What's different?
Kijai provided a great start, and I was able to build on that. It turns out - the 3d modeling part comes quickly and easily. The difficult part is texturing properly. I took a long time finding the correct camera flows, and how to upscale the textures so when they are painted on, they look significantly better.
Reactor: allows a character to keep facial features.
Ultimate SD Upscale - This allows me to push the texture to its limits while not being too liberal with changes
General upscaling - larger sizes are painted on, which allows a much tighter process.
More Camera Angles - The way it works is multiple camera angles are selected, then it bakes them all in. I went in and added a few more, and optimized the rotation around it to what I thought achieved the best results.
Granular approach - you can control how many vertices are on the object, and how big the texture turns out to be. a high res object is 3-4 MB and a low res object is more like 1MB. This means things like game assets can be kept small if necessary.
Extra Shenanigans - Mapping the camera angles is very VRAM intensive, so it's difficult to get a high resolution. However, I was able to split the camera into parts, allowing larger resolutions, and thus more accurate texturing.
PC specs - GTX 4090 with 64 GB RAM
Time to generate - About 20 minutes, but significantly less if you decide not to upscale.
Description
Cleaned up ComfyUI and added notes and FAQ
FAQ
Comments (12)
When loading your workflow the node Int is missing. Just replace with a new Int node? what resolution should be put in?
Yes, just replace it. I have it set to 1440 - it will resize the image to 1440x1440. Feel free to experiment - bigger is better, but bigger also takes longer to process.
I keep getting missing node type: ReActorFaceSwap error with this. Do you know how to fix?
Dunno when it happened, but that author's repository has been disabled on GitHub. That might have something to do with it.
I heard a rumor, and it appears to be true.
https://github.com/Gourieff/ComfyUI-ReActor
It has indeed been disabled - by GitHub - because reactor face swap repo was supporting nsfw face swaps. the new version has a nsfw filter on it.
You can very likely bypass the nsfw filter easily - he applied this to comply with regulations, not because he wanted to.
Please let me know if someone gets a fix for this.
How do you fix Hy3DRenderMultiView
No module named 'custom_rasterizer', i been fighting with this since i starte 2 days ago with this, and i read in blogs and i couldnt find answer, did you have this problem?
Hey there! sorry I took so long to respond! Actually, I hit this same problem. here was a fix that worked for me:
- You need to do a pip install . using the python.exe in your venv.
https://github.com/kijai/ComfyUI-Hunyuan3DWrapper/issues/14#issuecomment-2649695458
While the other answer can/could help you, i spent almost a whole day, as i had the same problem. I don't know what fixed it for me, or what combination, but i made it work. If you are lucky, just by matching your cuda and pytorch (And Python) version to the installation it will work, but even that i had one crucial missing part on my system, and it was Visual Build tools, with "Desktop Developement with C++" as well. If you try to compile your custom_rasterizer yourself, as the installation mentions for this problem, you may run into also that it misses something. There the tutorial states two folders you need to download and insert to the python_embeded folder (I tried both standalone and traditional installation. At the end portable worked, but steps does not really change, except you need venv, but maybe portable is more foolproof). I did that too, but the missing folder/file/thing still caused compiling to fail. Downloading the VS tools, and adding Desktop Development C++ were sure the last if not the only missing thing for me.
I keep running into a "Out of Memory" error when it comes to the Multiview node. Any advice?
Looks like a great workflow. but sadly with this as with 90% of other Hy3d workflows without the templates in comfyui desktop i get stuck on custom rasterizer. cant build it. installing doesnt help since its fro a different python usually or a different pytorch. il have to come back to this one at some point. its just. ahh so anoying. anyways thanks for this workflow tho <3
Sorry to hear that!! The stars must align with Triton, Cuda version, sage attention, and python version in order to make it work. Word is, its a lot easier on WSL. Maybe you could give that a shot?