For each model, see the version details for more information.
All the models here are experimental; they often won't work perfectly, the prompt will often need improvement, the input will often need alteration, or the parameters will need adjusting. Therefore, don't treat these models as something that should work perfectly right away for everything. They will only cease to be experimental when we discover a better way to do this.
Description
These are experimental versions; this is not v2, a continuation of v1. These two new versions are experiments I've been doing over the last month. The models are far from perfect, so please keep that in mind. In terms of motion transfer, it doesn't work well with fast-paced videos or scene cuts, so it's considered an experiment.
Read the README on HuggingFace Alissonerdx/EditAnything · Hugging Face
FAQ
Comments (44)
hi :) where can I find the workflow for the new "v0.1 - edit + motion transfer" version?
https://huggingface.co/Alissonerdx/EditAnything but remember this is not v2, these are experimental versions.
are the [ and ] part of the syntax? Or are they just there to indicate that's what we need to change in the prompt?
only indicative
Not working, i am getting just removed object every time.
Could you send a screenshot of your workflow?
Which of the experimental models are you testing?
@NRDX was testing old workflow from huggingface.
Your workflow have missing nodes.
LTXV Resize Reference By Mask is missing, i am updated BFS custom node its still missing.
@velanteg It's strange that nobody said anything about this. Could you check if you're using version 1.3 of the custom node?
Awesome material and research, felicidades Aliss!
hahah Thanks David
I used the provided workflow in huggingface but I keep getting the same output as the input for the reference video to video. Also, after a the second try I get this error: NotImplementedError: "mul_cpu_reduced_float" not implemented for 'Float8_e4m3fn" but not sure if this is from my end. I used the vanilla 1.1 as well as eros, both getting the same error
It's probably a model you're loading incorrectly, or even some incompatibility with your GPU, etc.
Happened the same two things:
Output video = to first one
On second run only "mul_cpu_reduced_float" not implemented for 'Float8_e4m3fn", but not on fresh run from launch
If I restart comfyUI I don't get error until second run again.
@AiLEarner Could some recent change in ComfyUI have broken something? I really don't know what it could be; if you're using the same base model I used, you shouldn't have this problem.
These are the kinds of questions LLMs excel at answering. When you have float activated the tensors become very light and drift off into space. You need to add weights so they can sink down to where the model lives. Comfy tries to make incompatible models live in harmony, but nonhuman nature most always prevails. They must complement each other, intertwine and mesh to complete the cycle of slop. It's science. Don't try to quantify it, accept it.
@NRDX i think it's Downloads/anime2real_01500_128.safetensors - your lora - I managed to find and download all else but this
I have the same issue. Output video is the same as input.
Everything works great. Thank you. You've done a wonderful job. Keep up the good work.
which prompt driven only ? it will do image to video ?
might be my favourite lora on this whole site!
Looks cool, can the lora be used on its own or must it be in your custom node setup
gods work.
Incredible work, I was using the workflow for adding a reference image but I was doing it with the first frame modified image, it was giving me great results somehow! Then I noticed the lora I had selected was not the correct one, turned on the correct one, turned off the module, and it worked even better but the identity of the faces got affected for some reason and that wasnt happening when using the reference image lora with the module
Will love to see this evolve more, sent you a coffee, thanks
Hey Lucas, first of all, thank you very much for the coffee. You were using the experimental rv2v with modules and then switched to rv2v without the modules, and you say the facial consistency wasn't the same? If that's the case, the problem is that the reference model is very experimental; it's not easy to make it work accurately. Normally, the latent position needs to be negative and not provide the image as a first frame (position 0). In the case of rv2v, I trained it focusing on a separate latent for the reference, which is why I created a node to do this automatically, but anyway, it doesn't work well. These are totally experimental models that I've been testing and training for about two months. You can try mix the loras, motion transfer + rv2v, I have a lot of experiments to do in the future.
We all know why there are no example videos from people by the community lol
Where in your workflows is the CFG? Is it hidden in one of the minimized connector nodes? I see no CFG anywhere? Having extremely limited success with this one, it seems promising but no matter how I prompt it basically either kinda works or flat out returns noise for faces, won't follow movement properly, etc. I have had a couple successful outputs but it then just goes back to failing to follow the prompting?
I suck at this stuff so don't take this as any kind of criticism please, thanks for your contributions to everyone here. Keep up the good work. ✌️
The value is passed through the subgraph node, and that value goes to CFGGuider
@NRDX I'm using the ltx23_edit_anything_with_ref_or_motion_transfer_v0.1 workflow, forgive my ignorance, I found the CFG... lol. Thank you for your response.
Sadly can't get it working with the workflow from huggingface. Using the "edit_anything_with_ref_or_motion_transfer_v0.1.json" but the process stops after it removes the background from the ref image and won't go any further.
Had great success with v1.0 though
update your nodes, mine was doing the same thing until I updated the node referenced in the error output, I forget which one but it should be in the console.
Take a look at the node inside the subgraph; it's possible that my looping sampler node has a property that's set to NaN because the custom node was updated but the workflow wasn't. If that's the case, I need to fix it.
@NRDX Thanks for the response. Got it working after switching the samplers, but generating with this is surprisingly slow. I've got a rtx 4080 16 gb card and 64 gb RAM and with EditAnything v1.0, I could edit an 11 seconds video in roughly 10 minutes. Same video, but with the v0.1 edit+motion took me almost 2 hours. Take it 16 gb VRAM might not cut it with this workflow
@Khevlar That's because it's v2v; the model is viewing frames from both videos simultaneously.
does it allow nsfw ?
use it with eros and it does :)
@ElGordoAI eros? can u link it? is it a lora or checkpoint?
@ElGordoAI I'm using 22b-distilled Q6 GGUF and not having any issues
so I did try 'remove cloth' for a scientific purpose but most of time it leaves sleeves. u got a solution for this?
Lol another lora? Plenty of them
mix it with eros or sulphur lora
@ElGordoAI I thought those are checkpoints
I can't get it to work. Trying replace head prompt used in example vid and the output is same as original vid. Using the workflow on alissonerx page. Maybe i'm using the wrong workflow, can anyone point me in the right direction??
Check if the Lora was selected correctly