Sign In

MovieStillXL - by Via

138
1.3k
6.7k
19
Updated: Aug 7, 2023
stylefilmgrainymovie stillsdxl
Verified:
SafeTensor
Type
LoRA
Stats
944
6,704
Reviews
Published
Aug 7, 2023
Base Model
SDXL 1.0
Training
Steps: 1,860
Epochs: 30
Usage Tips
Clip Skip: 1
Trigger Words
Movie Still
over the shoulder
in focus
action shot
Hash
AutoV2
CDF352508C
default creator card background decoration
viakole's Avatar
viakole

Movie Still XL V2!!

Please consider leaving a rating if you like it!

Movie Still XL is made for SDXL and its intended purpose was to add variety and coherence on the already great film and movie still images that are already integrated into SDXL. I am trying to reduce the need to prompt long film style prompts to achieve true movie like generations and with this lora it seems to work quite well.

This model was trained on 1344x768 images. The effects of other aspect ratios may be poor.

V2 NOTE: This model may produce green and pink tinting, if you are having trouble try this negative

'(blurry:1.4), (noisy:1.5) amputated, mutated, ugly, bad image, noisy, crispy, artifacts, tinted green, tinted pink'

What are the potential benefits from this lora?

  • Better "movie like" highlights

  • Better coherence "straighter straights, more consistency" (text and buildings are a little better)

  • Adds a filmic and movie like "Tone" to the final image

What are the potential downsides to this LoRa?

  • This lora can introduce a blur when it is turned up passed 1.0 strength

  • Sense this is still in early stage of development my dataset is still rather small so keep that in mind. This is basically an alpha version. Feel free to comment what else it needs!

  • Some images come out very grainy while some don't, results may vary

What are some of the trigger words?

  • Movie Still - the main trigger word to use the LoRa

  • over the shoulder shot - commonly used with a power of 1.5 to trigger that specific shot

  • 2 people and conversation - increased chance for there to be 2 people in a conversation

  • side view - increased chance to make the person side view probably mixed with 2 people

  • In Focus - Adds detail to background and makes it more in focus. maybe..

  • Action shot - adds action! this is already in sdxl but its also added to the lora itself.

That being said, again this is still in early development, this is my first published LoRa and I hope everyone who downloads likes it! I will be posting updated versions if I see an improvement as the dataset grows.

Recommended Usage:

Sytans SDXL v1.0 workflow in ComfyUI:

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

Best output:

The best output that ive got was from not using the refiner. I run the base model twice with the lora on >>> upscale 4x_NMKD superscale >>> downscale to a reasonable size >>> Do another 10 steps on the base with the upscaled image.



Strength: 0.85 to 1 ... higher than 1.0 can cause blurring or artifacts. use as you wish. may lead to some creepy stuff..

Thank you to Caith and masslevel for helping me with this LoRa!