3D, semi-realistic style LORA for Wan 2.2
Trained on 334 images, Joy-caption for captioning.
Triggers are nns3m1, digital art, anime style
Description
for Low noise model
FAQ
Comments (10)
Can this help animate 2.5D images with i2v? Have you tried? Thanks
I haven't tried with I2V, I mostly do T2V.
@NNAI No problem thank you.
What checkpoint are you using and what weights do you use for this LoRA? I like the looks but can't find more details in any of the images.
I'm not able to get anything anywhere close with this applied. Things become foggy, smokey, or blurry and that disapears if i drop this LoRA using otherwise the same generation settings.
I'm using Q8 GGUF models but the f8_scaled model works too, with 4steps Lightning LORA. euler/simple, CFG 1, 8 steps (4 + 4), LORA weight on 1.
If you download my sample video you can drag and drop it into comfyui, it should load the workflow from the video. I recommend using right click menu/save video option, the download button at the top not always work correctly.
@NNAI thanks, I was using 4 steps (2+2) which had previously been good for me, using checkpoints with Lightning baked in. I was looking through the individual posts because the first vid I tried downloading didn't appear to have the workflow. I'll give the other download option a shot.
Whelp, that helped find what was broken.
Changed checkpoints to test this, assuming the old one was inserting too much influence, and in the process broke one of my clip nodes (it's too bad the correct CLIP and VAE isn't baked into checkpoints lately). Seems that the new checkpoint needed the extra-steps as well.
Now to see what I need to tune to get the intended results, as adding "nns3m1" isn't resulting in much change to the style.
@Snowsexx32 I prefer the 250928 version of the lightning LORA most of the time, depending on the scene 0.7-1.0 strength for both models, lower strength can help with darker scenes. if too low movement starts to break.
The older lightx2v_v2 LORA better if you want dark scenes, but sometimes it has weaker prompt following and less influence of other LORAs, also I set the strength much higher to 2-2.5 on the high model for that LORA.
i m not able to fix nodes , tried everything from manual installations many times as well, is there something i m missing or is it a custom nodes?
NNVideoResolutionV2
NNAnyBoolSwitchTyped
NNSoftClearCache
NNWildcardProcessorin subgraph 'Prompt'
NNAnySwitch2in subgraph 'Steps'
NNImageCASharpeningin subgraph 'Post process'
NNContrastAdjust
All the nodes starting with NN are my nodes and not published, most of it are modified from another packages for my preferences or I couldn't make the original work as I wanted. For the LORA to work they aren't neccessary, You can use any workflow.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.