ComfyUI txt2img Upscaler Workflow
I wanted to share the workflow I am currently using. Also would like to know if anyone can give me tips and tricks about improving my workflow.. as expecting the 4x Upscaler get bit popular, I liked to showcase which nodes it needs. would like to know if you have any suggestion over the parameter settings, let's meet in comment section!
How to use: download zip file and extract all.
ComfyUI/
└── user/
├── default
├── workflows
└── 4xZIT.json
now you can restart comfyui and find it in your workflows.
SUGGESTION: using dpmpp_2m_sde and mode type Linear will adjust/fix your images
default: mode_type: "none" don't changes your image as final output.
Description
General Workflow for Z Image Turbo + Upscale
FAQ
Comments (9)
I have the same RTX 4070 and 64GB of RAM as you, but these upscale models never work for me. Your workflow, as well as others using these upscalers, always pushes my VRAM to 99%, and then my computer freezes. I generated an image at 832x1248 and then upscaled it to 1248x1992. I already generate my images natively at 1248x1672, and then they go through my effects and several upscalers, and I have no problems with that. My VRAM then peaks at a maximum of 90%. And I'm using an FP8 model, and even that can't handle your workflow. But otherwise, it's a nice little workflow that works, just not for me. I uploaded 3 images; the lowest resolution one is from your workflow with 1.5 upscaling, and that took me almost 5 minutes. The other two larger ones are from my upscalers, which take me about 3-5 minutes depending on what I turn on. The images then go through 2 samplers, 6 upscalers, 12 detailers, and over 20 photo effects.
%99 capacity is completely normal, it's taking avarage 100 seconds with 2.0 upscale for me. There are way to optimize your GPU/CPU shared usages. you can dm me if you still struggling to generate with upscaler
@loonalone My SeedVR2 upscaler only needs about 50 seconds per image, and I find the quality much better in direct comparison. My workflow involves many models: SAM, Grounding Dino, Depth Anything, QWENVL, SDXL, SeedVR, and then several upscale-by-model nodes. They all run in a chain with a maximum of 90% VRAM utilization. Since I generate natively at 2 megapixels, the image then goes to the upscaler and from there to all the photo effects, where numerous masks are created to mask specific areas of the image so that effects are only applied there. Therefore, these 4x upscalers are unusable for me because they throw an out-of-memory error. But if you simply want to take pictures without using any complex nodes, your workflow is perfect for that.
@nedolo77 I would love to try your workflow, you may be right that mine is not the best, it's just simple and works out.. would you mind sharing it? or, where can I find it, curious
why not considering seedvr2 for upscaling?
When you denoise image for upscaling it's generating new pixels with training data upscaler has. This upscaler fits more of the ZIT's style with less training images, giving smoother upscale compare to seedvr2 (no change) better than other high gpu demanding upscalers, it's effecting on the consuming time as well.. But I had prefer Seedvr2 if I am generating video honestly, it's just.. overkill for images 😅
@loonalone alright, sold. imma try this later tho. thanks
@loonalone Try generating a native image at 2048 resolution; the RTX 4070 can easily handle that. Then, use SeedVR and set its resolution to 2560 and its maximum resolution to 4096. In an image comparison, you'll see that SeedVR only sharpens the image and doesn't add much detail. However, if you generate an image at 1216x832 or similar, you'll have far fewer details, which the upscaler will then redraw. Like I said, give it a try; it won't take any longer than using your upscaler.
using using dpmpp_2m_sde and mode type Linear gives a super realistic touch to skin, thanks much!