Example images are made with a variety of checkpoints and prompts. Nothing unusual.
No trigger word
Model only (no text/CLiP)
Adjust strength to your own taste and what's in your chain
Will work better with some checkpoints than others
Hi-Res tends to revert the characteristics of LoRA on upscale
Let me know if you know how to improve performance with hi-res! Do I need to train in layers or mid layer?
Recommend checkpoints
Photorealistic checkpoints
Juggernaut | Analog Madness | Photon | EpicRealism | CineDiffusion
Description
FAQ
Comments (7)
Maybe it can work as a style, but it looks nothing like the film stock you're referring to.
How so? The colors look pretty close in some of the examples. I'm still working on improving the strength and generalization, though.
@motay41477317 I mean, I shot a bunch of film before, I can usually tell when I watch a movie if it's a real film shot, a reprint or just a grade. This, to me, really misses the mark of what film actually looks like: besides the color, it doesn't exhibit any halation anywhere, for example; while it's not necessary, it shouldn't be 0% of it either. Film scans do look way, way different. Especially when you're claiming a name of a leading film stock (like Vision3). People have been shooting these for decades, you have hours of source material. All these look like just a bias to greenery and commercial photography, they do look nice, but they don't resemble film scans to me at all.
https://www.flickr.com/search/?text=fuji+pro+400h&view_all=1&sort=date-posted-desc
Again, it's nice if it's directing you in a direction you like or people like, nothing wrong with that, it's the nature of the beast, of course. :) LORAs are cool. So no offense meant.
@motay41477317 I guess I didn't tag the reply properly, there you go.
@Vospi Thanks for the feedback! I can see what you're saying. Some of the images show halation and a flat profile, albeit too moderate. However, many of the images are lacking the traits you expect. I've been thinking about your feedback and troubleshooting a bit what may be the cause. It's very easy to go down blind alleys when trying to optimize training of LoRAs given the number of parameters you deal with. Sometimes you focus on improving one metric over time, and you may not notice another metric slowly degrading in the process.
That being said, I think I may have pinpointed the biggest issue. It seems the weight decay I have been using (0.01-0.03) is too big for the size of my datasets (250-1200). The results I'm getting have improved a lot by turning it off alltogether. I will do some more optimizing of settings and publish much improved versions of my LoRAs over the next weeks.
@Vospi Does this look better? https://civitai.com/posts/3257269
I'm planning to improve it some more, but I think it's looking a little more realistic now.
@motay41477317 I'm very happy I didn't offend you; 0.5 look interesting in general. I (kinda) understand the black box nature of what you're dealing with, good luck. :)