This LoRA is designed to capture the distinct look and feel of early 80s fantasy movies.
The trigger words are ArsMovieStill, 80s Fantasy Movie Still
Z-IMAGE UPDATE:
V2 dataset officially became my go-to test dataset when training a model for the first time.
Experimental!
Default ComfyUI LoRA Loader Reports Errors, but actually applys the LoRA.
Recommended usage settings:
Your favourite spot.
Listen to Castle Rat as background.
Enjoy!
MJ7 UPDATE:
Building each dataset is a joy, so with the release of MidJourney 7, I naturally started this one again from scratch.
This version stays true to the original aesthetic but with improved fidelity and an even stronger 80s vibe.
It’s trained on a fresh, carefully curated set of ~250 images, created by referencing the best results from the previous versions.
V3 UPDATE:
The training dataset has been further expanded to 1,000 curated images, pushing it to the limits of Civit’s on-site training capabilities.
This update include broader selection of fantasy environments, mythical creatures, and epic scenes, enhancing the fidelity to the iconic 80s fantasy aesthetic.
V2 UPDATE:
Added to the dataset a huge amount of Fantasy Characters, beasts and scenes.
Use v1 for stronger effects and v2/v3 for better versatility.
Description
This is Trained on 1 Frame with the V2 Dataset.
Adding it if anyone uses T2I Workflows that use High and Low Wan2.2.
It can be also used as T2V, but you want to crank up the strength to 1.5 for visible effects, which can fry the video.
Produces quite decent results as a I2V, if you add as a starting point an Image produced with just the Low noise LoRA in a T2I workflow.
Have fun.
FAQ
Comments (17)
I love projects like these because they also show how the models themselves perform. Those WAN examples are incredible.
WAN 2.2 looks really good!
Yeah, I came here to say that it definitely has a strong and unique look, even compared to all the other versions of this lora.
who said that wan 3 is coming out soon?
There were a bunch of rumors that 2 new Big Models are coming from China before the end of the month - TI2V model and new NanoBanana competitor .
@ArsMachina wan 2.2 literally just came out so it would be surprising if wan 3 came out so soon. im prepping a dataset for wan training now i wonder if i shouldnt waste my money, and wait a bit now lol. what are your discoveries on training wan vs other models? some say its difficult to train others say its easy. im sure that just means they had different goals. like you, i want to train a style though so im more interested in that. is it harder to train? the images for this lora look great btw, wan just has such great background detail that makes the images feel bigger and more powerful
@cutetodeath78409597
Might not be Wan 3, but some other Chinese model then :) They come pretty frequently recently and the rumors seems to be correct so far.
The Training is MUCH easier than Qwen and MUCH harder than Flux , so everybody seems to be correct :) My dataset was 546 Images and the Model was OK after 1 pass on each image at 550 steps. I did 2 passes , just so we have better coherence.
I used AI Toolkit With Default (Low VRAM) Setting Q8 both for Model and Encoder - first training low, then High with LR 1e-4.
With this setting the model fits perfectly inside 5090 and runs ~2-3s/it (step).
@ArsMachina i saw some rumors on reddit that its not wan 3 but wan 2.5 lmao. thank you for all the info you are awesome. wan 2.2 STYLE loras are a rarity here, thx for your work, love this model for image gen so im starving for some good ass loras like yours 🙏🙏🙏
please more chroma updates for the rest of the models. we have so few chroma loras.
I think I will include Chroma in my 5 "Default" models I create with new datasets. So far I like it alot :)
@ArsMachina U use the on-site trainer thingy right? Does that use the yellow buzz stuff or the blue one?
@TijuanaSlumlord Yes, the Chroma LoRA is trained on-site. It first uses all your Blue buzz then uses Yellow one. I do not recomend increasing the setting above 1000 Buzz per training as almost 100% you are overtraining .
@ArsMachina I see. Considering i used your Loras a lot and i know how good they work, i hope this helps a bit. I do everything local, so i guess it makes more sense giving it to u... and as a thanks.
@TijuanaSlumlord So you make the Generious Donations? Thank you :) I really can't see who made the contribution to a model, I just see date and amount.
@ArsMachina Most likely just a bug, pretty common with Civit nowadays.
Chroma LORAs are rapidly uploaded now. This is our Pony moment
Great job everyone!
@ailu91 yeah, its my new default model and has replaced pony/il, the gguf version is pretty good, upscaling with kontext is my go to worflow.