A lora to celebrate the science fiction movies and shows of the 1950s. It's fairly narrowly trained, doing well with space suits, bulky robots and moon-like landscapes.
All images/videos in showcase contain ComfyUI workflows.
Hunyuan 1.5:
First try. All examples done without lightx, 20 steps.
Wan 2.2:
The Low noise lora has absolutely most effect for style. High noise still affects the outcome, but to a much less degree.
For the Wan showcase, I use 25 steps: 15/10. 0.8 strength, and dpm++_sde scheduler in ComfyUI. I haven't had a lot of time for testing, but it might be that it loses strength as you go up in frames.
I use Q4.gguf for High noise, and Q6 for Low noise, for no particular reason. No block swapping.
Wan lora trained with diffusion-pipe.
Description
I've reused a lot of prompts from the Hunyuan version for you to compare.
I'm a bit unsure about this version. Does it need more training? Sometimes it looks great. Sometimes the style is right, but the quality is too good for being 50s. Like Wan wants to "fix" it. I'm experimenting a bit by removing the negative prompt.
FAQ
Comments (7)
What do you think about the Wan 2.2 version compared to the Hunyuan version? Do you like the results?
I think that Hunyuan in general feels more consistent in style. I feel like Wan tries to improve the result, so the style is kind of the same, but the image quality better and more modern. Same goes for camera movements. But it could also be that the lora is undertrained. I might do another 10 epochs and see.
Edit: It could be a number of other causes as well, most likely skill issue on my part.
neph1 If you do more training, I'm very interested in the results. I'm working on something similar for 80s Japanese sentai TV shows. Specifically the effects, but I also want to keep the visual look like yours seeks to do. I love the look of those old shows so much. All on film, and with hand painted effects. I really hope there's a way to pull it back from looking too modern.
I do like your results on their own terms though, even if it doesn't quite go to where you want.
Jellai Will do. I realized it might help to lower the resolution. I had forgotten this training data had lower res than the fantasy one, but kept the generation settings. 352p seems to keep the style better.
neph1 I found some other training instructions for Wan 2.2 that said to use 256x256 video on High, and higher resolution images on Low, to target details and motion. The 352p training data you used, was it images or video? And did you change the data for low and high?
Jellai Only images. And it seems High has little impact (but some) for style loras. Used same. I'm still learning, as well :)
Really like this !!!! (wan 2.2 version)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.