ImageUpscaleFlow is a dedicated workflow designed to upscale images using SeedVR2, currently the most powerful and freely available upscaler for ComfyUI. The example shown in the gallery demonstrates an impressive upscale of a 434 × 652 pixel image to a stunning 2484 × 3732 pixels. The remarkable quality of the result truly speaks for itself!
💡If you installed the SeedVR2 node before 2025-12-06, update it using the ComfyUI Manager. The latest version significantly reduces artifacts, especially when using the 7b_sharp model.
Description
FAQ
Comments (22)
So far this is the simplest version of this workflow and it works really well. Good job.
Finally, a SeedVR2 workflow that works without having to manually install nodes or anything, this is perfect. The instructions were very clear too. No problem on 16gb VRAM.
Which configuration do you use? I guess my default settings (optimized for 32 GB VRAM) do not work.
@ai839 just enabled tiling and block swapping at 18 or 27 and its good to go. less than 30 seconds for max size.
@slimd474414 1024 as tile size? It would be interesting to know what the maximum tile size can be. For example, 1536 or 2048.
@ai839 yep, kept all values at default except for block size
@ai839 My Tile size 1024 enabled 2 blocks block swapping 32 ----- didn't worked for me.
I have 8gb vram and 32gb ram. using 7b but on other workflow I can generate 1536 + 3840 resolution but still working.
@aneebtheartist126 Lower the target megapixels. The default is extremly high. Maybe start with 3 and then go upwards until your limit.
This!
Thank you sir, this is an example workflow done the correct way. Just dl the nodes read instructions and good to go. It is so refreshing to see a well structured workflow, with proper instructions.
wow that so fast and better what i expected.
do you have the video version? got bunch of wan videos in 480p lol
No, but the official tutorial video explains it very well: https://www.youtube.com/watch?v=MBtWYXq_r60
@ai839 not worth it bro, upscaling above 10min for 5sec video is waste of time.
i did try individual images sequence it took me 10min for 80frame, but of course it's lacks the consistency when it's actually combined into video. worth a try tho
change model to 7b_Q4_K.gguf if you have small vram
Why your workflow keep running on 95% Vram and 79% Ram even I refresh everything with free model and node cache. you workflow is locked too. What's going on?
Follow the instructions to understand how to manage the memory consumption. Especially the linked YouTube video from the author of SeedVR2 is helpful. The default values in the workflow are for high-end hardware like 32 GB VRAM and 96 GB RAM.
Is there a way to load multiple images?
You can batch-process all images in a folder by iterating over them automatically. Add a Primitive node (right-click → Add node → utils → Primitive) and connect it to the left input of the image field on the Load Image node. In the Primitive node, set control after generate to increment so it steps through each file in the directory. Now start the workflow with Run (On Change), and every image in the folder will be upscaled one after another.
@ai839 thanks for that :)
Out of curiosity, would installing sageattention (and is prequisites such as triton) speed up SeedVR2 single image upscaling in a meaningful way? It's such an involved process I'd rather not unless I get substantial gains. Plus, the 3b models are already very fast on my setup, so it would mostly affect 7b models, which frankly didn't seem to improve very much over 3b.
I tested SageAttention 2 for you, but I didn't notice any remarkable speed improvement (more than 8 megapixels resolution, no tiling, 7B model). I'm using an RTX 5090, so it's already fast without SageAttention (aprox. 20 sec). It might be more relevant for video upscaling or slower hardware.
@ai839 I see, thanks. I'll skip that then, since I'm not planning on doing video generation. While my setup probably could handle it (RTX 5060) I'm pretty sure even a 10s clip at 480p would take a good while to render.