Support my work or request a custom lora here! Even 1$ can make a big difference ☕Join patreon for exclusive contents!
UPDT V2.0
As requested by HeeroAmada the pony version for yanderechan is here!
Image are generated with 512x1024 res and then upscaled, the result is great! Like every Pony models and from my experience do not forget to use:
Positive prompts:
score_9, score_8_up, score_7_up, just describe what you want, tag1, tag2,etc...Negative prompts:
score_6_up, score_5_up, score_4_up, etc...V1.0
Hi everyone :D, this time i've got a surprise :D, i finally made an animation with Animatediff extension! It took me a while and the use of another ckeckpoint that i will add below.
Yandere Chan"A Girl who came from another world. She has a secret and a mysterious aspect to her personality". This Lora was created following this easy Guide and trained on Meinamix11
Main tag: yanderechan,black hair,silver/red eyes,ponytail.
Weight: 0.6 - 1.0
Description
FAQ
Comments (26)
A-W-E-S-O-M-E!!! ❤️ How long does it take to make a little vid like that?
Thank you so much :D, it depends. In this case there are 5 different 2 seconds animation, i spent a lot of time researching a good one by generating at least 10 for each, then choosing the best one. Since I had to generate the model first , im gonna say 2 days ;)
Are you willing to share your workflow here? :)
Sure! If any info is needed, just ask ;)
---IMAGE GENERATION---
1 - Find a good prompt/idea for generation
2 - Using Darksushi 2.5D model set:
Sampling DPM++ SDE Karras, Step 20, CFG scale 7, desired/capable resolution baste on your GPU. this case 512x768
3 - Set animatediff:
Model mm_sd_v15_v2.ckpt , Number of frame 16, FPS 8, Frame interpolation Off.
4 - Found the desired animation, it will generate 16 images save them and put in a folder.
---IMAGE UPSCALE---
5 - Trough the img2img tab -- batch set:
the same prompt and seed generated previously
imput directory: folders with 16 img
output: prefererd folder
Sampling DPM++ SDE Karras, Step 20, CFG scale 0.35
Resize by 2
6 - Enable Adetailer: face_yolov8n.pt (in the prompt i added the lora, thename of the character and eyes color)
7 - Enable ControlNet,pixel perfect,Tile/Blur and Tile_resample with control_v11f1e_sd15_tile, controlnet is more important.
8 - Press Generate, it will upscale all the img inside the input folder.
---ANIMATION COMPOSING---
9 - Open a CMD inside the folder where the rife-ncnn-vulkan.exe is located and launch the command:
"rife-ncnn-vulkan.exe -i <your_folder_path_with_16_img> -o <your_output_folder>" to generate from 16 img to 32 and so until reach 128 (and deleting the final 8), so keep 120 imgs
10 - use ffmpeg commands "ffmpeg -framerate 60 -i <your_folder_path_with_120_img>/%08d.png -crf 18 -c:v libx264 -pix_fmt yuv420p output.mp4"
Enjoy!
@Mario1964 Thanks!
Maybe this is obvious and I missed it - are you using A1111, Comfy, SDnext?
@vokar28 A1111 ;)
@Mario1964 Awesome. I will give it a try! I was having issues with a1111 freezing, but I was trying to do hires and adetailer at the same time as animatediff, so maybe it was just overloading.
One other thing I noticed last time I tried was that a1111 (and comfy) was generating one big 4 by 4 grid, rather than 16 individual images. Is there a specific setting that controls that?
@vokar28 those are both GPU demanding, It's always suggested to do them separately (hires e animatediff).
Apart from installing the extension,disable in settings Optimization - Batch cond/uncond , i didn't change anything.Maybe try a clean install :)
@Mario1964 Strange. I'm getting an error: "OverflowError: Python int too large to convert to C long" when trying to do the img2img step. Any insights into that?
@vokar28 You mean doing the batch generation part. Mhhh i've got "an error" in the past, but updating the extensions it solved. You could try to doing it manually and insert every single img into the img2img to upscale them :(
@Mario1964 Yes, when I try to to do the batch part. I've updated everything, so I don't know what's wrong. I may try a clean install, or just do the process manually. Either way, even the non-upscaled animations are looking better with this method! :D
@vokar28 My apologies if can't help more, hope to see some your animation ;)
@Mario1964 No problem! If I sort it out I'll post it back here. I'm tempted to try this all out in SD Next to see if I get the same issue.
For the rife-ncnn-vulkan section, I had to put quotes around my paths for it to recognize them. Just FYI!
Now rife doesn't seem to be doing anything other than display 4 lines of information and spit out the input frames back into the output folder with no new generations. Hmm... Interesting puzzle.
@vokar28 Inside the cmd,when writing the command you can also drag and drop your folder input and output without copy/pasting the path.
Steps are right, it will shows 4 lines and after a bit the process ends. You are facing a lot of problems that I didn’t get :/ you could try to use Flowframes, i didn’t try and don’t know how it works but it’s a solid alternative!
@Mario1964 I will keep doing some basic troubleshooting and see if I can get around it. Otherwise I'll give flowframes a try! Thanks
Please clarify for me if you can, RIFE is supposed to generate new frames, but since we're not specifying how many frames, how does it know how many to generate? (mine still doesn't want to generate new frames up to 128 like in the steps, not sure why!)
@vokar28 As posted in this Guide that i followed for those steps:
Open a command prompt and execute the following command, specifying your input and output directory paths:
rife-ncnn-vulkan.exe -i <input_directory_path> -o <output_directory_path>This command generates interpolated frames between every two frames, resulting in double the number of images in the output folder compared to the input.
I dont know more hahahahah
@Mario1964 Ohhh, so it's adding frames in between to increase the FPS and number of frames, not appending more to the end to increase the length. I guess :D
Woah! Technology!
Hello I looked at your tutorial and I've gotten to the point with the RIFE, I have absolutely no idea how to go from here, I got my images, I upscaled them etc. I downloaded RIFE,CUDA10.2 and all of the other dependecies and now I have no idea what to do next. There is no exact tutorial on the matter I just followed random tutorials for other things that use RIFE and I took a bite out of every tutorial and now I'm here. I looked over the github page for RIFE and I still can't figure it out. I dislike coding and I never understood python that much so maybe that's the issue I'm having as of right now.
Hi, there is 0 coding or hard steps to follow, it's a simple things :) I already updated the workflow under the animation btw;
Open a CMD inside the folder where the rife-ncnn-vulkan.exe is located and launch the command:
"rife-ncnn-vulkan.exe -i <your_folder_path_with_16_img> -o <your_output_folder>" to generate from 16 img to 32 and so until reach 128 (and deleting the final 8), so keep 120 imgs
if you still got problems, lmk :D
@Mario1964 Ah I got it now thank you very much !
For anyone reading this and having problems on the last part, you need to download RIFE and follow the instructions from Mario1964 and then you need to download the full version of ffmpeg and then extract it somwhere and go into the bin folder inside and from there open the cmd prompt through the folder, here "D:\ffmpeg-2023-11-20-git-e56d91f8a8-full_build\bin" and then follow Mario's instructions again, the output should be in the same folder where the 3 .exes are !
Thank you again, was fun working on this ngl !
Another question:
It seems like the output is different after adding the animatediff model into the equation. Is there a way to prevent the animation models from changing the image?