• I added a simple workflow adapted from default ref2va workflow. This lora is designed to restore enhance and improve low quality videos. It helps with upscaling and it also helps remove things like block artifacts, compression noise, and blur.
• This lora can also help extend your video as well. Simply make the duration of your final video longer than the short original and it will naturally extend it.
• Use lora Str 1.0, connect your video to the video reference and make sure your latent size ratio is the same ratio as your input video.
Restore this video to high visual quality by recovering fine detail and reducing pixelation, block artifacts, and compression noise while preserving the original scene, natural colors, and motion. Description
v1
FAQ
Comments (28)
nice. did a good job on old home movie
How does it compare to LMS?
https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI/blob/main/docs/lms.md
https://civitai.red/models/2923395/lms-a-little-more-sharpness
That one is really incredible at restoring old low res video
yes i tried it but i guess you can only use at strength of 0.5-0.6 more than that the character face changes even with ref2va
I have tried a little more sharpness and it is a good model, instead of making it very similar to that model I tried to focus a little more on different aspects like artifacts and compression instead of just sharpness
@Nugus your lora on strength of 1 actually holds the video.
@Nugus ah, nice, glad it focuses on those other aspects. I don't think LMS was intentionally starting out as trying to be a video enhancer at all, just happened to work out as one. I'll have to play with using these together. LMS doesn't really do anything with video reference inputs, only guidelines, so there's probably a nice balance to be found combining them since they work on separate piecies
@phazei thanks for bringing this one up. Its really good!
I added a workflow, https://civitai.red/api/download/models/3389308?fileId=3280520
can you add a default workflow and also does it mean that the lora will have a lock on the video file.
I added a workflow, https://civitai.red/api/download/models/3389308?fileId=3280520
my input output both had same res it didnt do anything
he mentioned to upscale video using upscalers. well I had good results
@chaiwala008215 As I see it for now:
1. Generate your video at low res (e.g.: 2:3 - 240p - 256x384)
2. Use your output as video reference with the lora and specified prompt
3. Manually increase resolution output maintaining the same ratio (e.g.: 2:3 - x2 = 512x768)
4. Run
5. Uspcale ->done
Works for me. I use it as second pass, generating first output at low res in less than 1min, reinject the output directly into a second pass sampler as video reference using the lora and the prompt with doubled dimensions, then upscale VSR x2 for the final output. Same total generation time as if I did it directly at the second pass resolution (manual x2), but without uspcaling defects with it.
The increase in res is something you manually set in your workflow like K0DEX was saying, set your latent size to 2x your original video size is a good starting point but it all depends on how big your original is.
@Nugus Nice job by the way! Have you tried using it on direct generation as other LoRAs to improve quality? I wonder if it works as well...
@K0DEX thanks, I haven't tested that, let us know if you try
can people post their results so we can see what this LoRA can do when properly used?
I added a workflow, https://civitai.red/api/download/models/3389308?fileId=3280520
can people post their results so we can see what this LoRA can do when properly used?
and how do u use this Lora in a workflow?
Go ahead and post, bud
@ploxjunkemail793 I have no idea how to use this lora sadly... anyways i tried out the LMS one and it's been really good so far.
@hatt2 you need to just connect your video, add the lora,apply the prompt OP provided and run it
where do I get the node MinimaxH3cropto32?
Works great! Only problem is it have tendency to change videos. Especially on 4,3 step turbo loras ....or at no turbo 20 steps! Best results on 8 step lora from lightx or larry.
The best way to upscale video I know right how.
I noticed this as well but I played with it a bit. Changing resolution gives the AI more space to add. Example I ran a 320x240 video into a 16x9 format. I added all kinds of stuff. Reran it in original resolution just to get better details and it stayed mostly the same. More are means more play time for the AI. I could just be reading into things to much. Ive been using rtx long to upscale after original video resolution has been altered.
Damn, I have 0 complaints and have even used it with the multishot chain workflow and have redone entire videos of 3 minutes and up. Rock solid bro!
Have you tested this LORA alongside the generation itself, and if it helps keep artifacts out of lower quality videos?