Remember to read the instruction!
Please use Pixel version, if you don't have good understanding on Latent version.
This will provide you
Crazzzzzzzzzzzzzzzzzzzzzzzzzzzzzy
Amazzzzzzzzzzzzzzzzzzzzzzzzzzing
Higggggggggggggggggggggggh
Quaaaaaaaaaaaaaaaaaaaaaaaality
Images!!!!
just try it!
remember use my custom node and modify the parameter carefully!
adding jc2 version. Remember to download my nodes
https://github.com/TTPlanetPig/Comfyui_TTP_Toolset
https://github.com/TTPlanetPig/Comfyui_JC2
check your Max VRAM before you run it :)
at least 10GB for nf4 and 20GB for BF16
Following is for Pixel version instruction:
Just enjoy, and try not to touch my setting if you are not sure what you are doing.
if you can run flux official properly, you can run this one
Following is for Latent Version instruction:
This is a workflow for my simple logic amazing upscale node for DIT model. it can be common use for Flux,Hunyuan,SD3 It can simple tile the initial image into pieces and then use image-interrogator to get each tile prompts for more accurate upscale process. The condition will be properly handled and the hallucination will be significantly eliminated.
you can download my newly build node here:
https://github.com/TTPlanetPig/Comfyui_TTP_Toolset ,please update to latest version
A updated version for Flux Hi-res fix!!!
adjust the denoise to control the difference and possible latent block issue.
You will be shocked by the detailed high-resolution photos!!!
Key parameter to adjust:
1. tile block numbers: recommended 2*3 or 3*2 or 3*3
2. overlap rate: recommended 0.15-0.2! Please update my node to latest version for this function.
I also used the node:
https://github.com/yolain/ComfyUI-Easy-Use
and
https://github.com/kijai/ComfyUI-Florence2
Description
key update!
FAQ
Comments (14)
Niu B Plus
ni tai niu B la!
Very interested approach, good work! I recommend in the next release to include the loras.
lora is supported in my tools，just insert in the middle
@ttplanet I tried three different uploads of the “lora,” the initial image is right, but the final upscaled image comes out ugly and blurry.
@superacf135 lora clip should be also included in the upscale clip for condition, did you do it correctly?
@ttplanet The lora node has two outputs: model and clip, I simply connected the nodes as they were in your workflow. Here a screenshot: https://paste.pics/a81f04790d929175634b4b16878dd91a
@superacf135 I will run a test soon as I am working on my training set now
呃 呃 啊 啊啊啊啊 ~老白 ，用力~啊 啊啊啊
Still from a movie, a photo of a young Japanese female explorer, with big wavy hair, huge sexy breasts, wearing a futuristic sci-fi cape and a futuristic sci-fi hat, walking in the desert, with sci-fi buildings vaguely visible in the undulating desert. A sign with the words "TTP Planet" stands on the mountain, with movie lights, shot from above
Niu B Plus
You have done a wonderful job overall, but unfortunately, I won’t be able to fully utilize it. It’s quite heavy for my system (RTX 4070 Ti Super). One image took me almost 7 minutes. Also, the flash ATN doesn’t work for me either. I have installed the SDPA. I need to find the right settings because I’m getting double images, blurs, halos, etc., which takes a lot of time to fix. It’s a shame because what you’ve created looks so good. :(
You can see others sample, it will work with proper setting. I even want to make a batch logic to make it even faster.
The key is monitor your VRAM make sure it does not over run, by the same time, I will assume Florence is a low resource consume model which can make prompt, you can replace it and input manually to further save the resource, please notice the process actually boost 1M image to 8M Directly, you can adjust it to lower size for example 4M
Remember to change the tile number from current 6-9 to maybe 4-6
Let me know if it works
请问 文本连锁这个节点的用处是什么? 他没有连接到任何地方. 可以删掉他吗?