List of changes in v2:
Added base model selector, from FP16 to Q2
Individual activator for SageAttetion and Torch Compile
Added “Image Style Loras” panel to change the style of the generated image. “Smartphone Snapshot Photo Reality” has been moved to this panel along with other style loras. The download links and recommended strength are available there.
Added option to select the total steps, with automatic calculation for the steps of each KSampler.
Added “Prompt variation helper” option to help get more variation in the result.
Added option to use VAE or Tiled VAE
The generated image is now upscaled to x2 by default.
New settings in KSamplers to prevent image defects (body elongation, duplication, etc.).
New image enhancement options, including Instagram filters.
Additional upscaling options to x2 or x8 (up to 30k resolution).
Hi!
I've customized a workflow to my liking with some interesting options and decided to share it.
Hope you like it.
Here are some details:
Ready for GGUF models and MultiGPU
Option to easily enable/disable basic Loras (Lightx2v, FusionX, Smartphone Photo Reality)
Option to enable/disable additional Loras (characters, motions)
Option to select a preset size or customize it manually
Option to add sharpness and grain
Option to enable Upscaling
Option to enable accelerators (Sage Attention + Toch Compile)
Descriptive text for each step
I used 2x3090Ti and the generation time at 1920x1080 is about 100 seconds.
* For the size presets you will need to copy the “custom_dimensions.json” file into /custom_nodes/comfyui-kjnodes/
If you encounter any problems or have any suggestions for improvement, please let me know.
Enjoy!
* This workflow works really great with character loras.
* The first images I uploaded got taken down, and I got a warning for using images of well-known people, so I re-uploaded images created with the WF without character loras.
Description
Hi,
I've been working for several days on v2 of the workflow
There are several new features that I hope you will like and find interesting.
This WF is more complex than the previous one, but I have tried to detail each step and explain the new options.
List of changes in v2:
Added base model selector, from FP16 to Q2
Individual activator for SageAttetion and Torch Compile
Added “Image Style Loras” panel to change the style of the generated image. “Smartphone Snapshot Photo Reality” has been moved to this panel along with other style loras. The download links and recommended strength are available there.
Added option to select the total steps, with automatic calculation for the steps of each KSampler.
Added “Prompt variation helper” option to help get more variation in the result.
Added option to use VAE or Tiled VAE
The generated image is now upscaled to x2 by default.
New settings in KSamplers to prevent image defects (body elongation, duplication, etc.).
New image enhancement options, including Instagram filters.
Additional upscaling options to x2 or x8 (up to 30k resolution).
The next version may use Qwen as the initial step, or have some image2image control... but for now I'm going to take a few days off after many hours of testing, lol
Enjoy!
FAQ
Comments (29)
cant seem to get the High/Low SNR section or node work. Despite manager saying i have all the nodes installed without issue, and having updated them all to be sure.
That node was created by me because I couldn't find one that did what I wanted. It was my first time creating a node and it took me a while, lol. Although its function is very basic.
In the download link, you'll find a TXT file with instructions on where to copy the necessary files, which are just a couple. The custom node goes inside the “custom-nodes” folder in the root directory. The other file is for the preset sizes and goes inside “custom-nodes/kjnodes.”
CaptainHarlock80 Thanks, i must have overlooked the instructions, i'll see if i can get it working now.
whitespider9999 You're welcome. No problemo, let me know if you have any problem.
CaptainHarlock80 question, in the ultra section of the VRAM/wan 2.2 checkpoints, the top one is I2V, was that intentional or a mistake?
CaptainHarlock80 And thanks for before. I should check instructions next time.
whitespider9999 Oops, it's a mistake.
CaptainHarlock80 all good, i'll get the t2v checkpoint.
whitespider9999 Great, enjoy!
CaptainHarlock80 when you update your workflow, a few little details
- cuda 1 is set, which means you probably had two nvidia gpus, this might confuse someone that has error messages saying how it went to the cpu, for people with 1 nvidia card, cuda 0 will generally be what they have to set. cuda 1 will default to cpu if its not present.
- The text in the vram/checkpoint is incorrect on a few areas, it also says "HIGH/HIGH on the ultra setting" not all the nodes have been correctly routed. Easily fixable, but might trip up a novice.
whitespider9999 Yep, I do indeed use 2 GPUs. I thought that if there was no CUDA 1, CUDA 0 would be selected automatically. I'll keep that in mind, thanks!
Yep, v2 is more complex than v1. I've tried to make it as simple as possible and include explanations, but if anyone wants something simpler, they can use v1.
Thanks for your suggestions!
CaptainHarlock80 yep. all good. Loving the results now it's all up and running
whitespider9999 Great !!!
From civitati: "This image has been blocked because it is has received a NSFW rating and we could not verify that it was generated using AI"
I guess that's good, LOL
Great work! Thank you!
I absolutely adore the female faces that WAN2.2 produces! Pretty, but not over the top glamour or fugly (whne using Loras to take of the gloss) like Flex. Naturally Beautiful.
Hi, this is great! Is there any way to possibly add a reference image to it?
I haven't tried that with the Wan2.2 version yet.
I saw that Fun came out, but I haven't tried it. I'm waiting for the VACE model for that.
You can use VAE Encoder to insert an image into the latent of LOW's KSampler (you don't need the HIGH one), just remember not to start at step 0 because if you do, it won't take the image into account at all.
i want just 720x1280, how i can do that
Sure, no problem. You can do it in two ways.
Go to Preset Size and choose 720x720 (look at the right side of the drop-down menu, because on the left side the size is rescaled to x2, meaning that the GPU will actually work as if it were 720x720, but the image you get will be 1440x1440.
Or you can disable Preset Size and manually set the resolution you want to use (it will also be upscaled to x2).
Sorry, I see now that you were referring to 720x1280. In this case, if you use Preset Size, select 1280x720 and enable the “invert” option. Or enter the size manually.
hello,which lora do you use?
Which one are you referring to exactly?
For Lightx2v, this one: "Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank32"
For FusionX, this one: "Wan2.1_T2V_14B_FusionX_LoRA"
Anyone have problem with WanVideoNAG?
Does anyone have the same blurry image as me? Even though I increased it 20 steps and turned off the acceleration loras, the picture is still blurry! I am using Wan2.2-T2V-A14B-HighNoise-Q5_K_M and Wan2.2-T2V-A14B-LowNoise-Q5_K_M.gguf
Ensure you install the samplers/schedulers used by the WF (https://github.com/ClownsharkBatwing/RES4LYF), or use others. Check your HIGH+LOW step settings.
@CaptainHarlock80 Thanks, I increased my steps and am quite satisfied with the results. Your workflow is great!
@traithanhnam90 I'm glad to hear you figured it out!
Between 8-12 steps for the 2 KSamplers should be enough.
BTW, you can download v3 of WF here with versions with and without MultiGPU: https://www.reddit.com/r/comfyui/comments/1mlvwh1/wan_22_text2image_custom_workflow_v2/
I have many nodes missing that cannot be installed thanks to the new update from ShittyUI, any idea?