This LoRA model captures the unique visual aesthetic of the Olympus D-450 Zoom digital camera, a popular 1.3MP point-and-shoot released in 1999.
Trained on a dataset of 1,100 photos over 4,000 steps using AI-Toolkit, this LoRA replicates the camera's effect, including its signature soft contrast, limited dynamic range, color fringing, lens distortion and a slight grain effect.
The model is perfect for adding a nostalgic, retro vibe to your AI-generated images, giving them that distinct early-digital texture while preserving natural light tones and details. Ideal for artwork requiring a warm, vintage look or for users who want to evoke the era of turn-of-the-century digital photography.
Trigger Word: olympusd450
Use this word in your prompt to activate the Olympus D-450 style.
Details:
Trained on: 1,100 photos (1280 x 960)
Training steps: 4,000
Software used: AI-Toolkit
Best suited for: Portraits, landscapes, and everyday scenes
Enjoy bringing a bit of early 2000s nostalgia to your creations with the Olympus D-450 Zoom Style LoRA!
Description
FAQ
Comments (19)
None of the images posted by others look as good as yours. Can you share a prompt?
Sure, the ones by @peterkickasspeter look great though! The key is to use a LoRA weight of at least 1.5, I actually prefer 1.7.
Here's the prompt of the first sample image:
olympusd450, taken at a party with flash, group photo with 3 people, 30 year-old man holding a can of beer, he wearing a red mario hat red T-shirt and blue suspenders. Next to him is another man person wearing a brown sweater and a furry wolf mask on his head costume, standing next to woman wearing a zombie nurse outfit with green and grey face make up, a red cross on her blood-stained white gown, flash photograph, Jack-O-Lantern Halloween Lamp in background, people in the background, detailed skin
Resolution: 1280x960
Seed: 805951353726032
Scheduler: sgm_uniform
Steps: 35
Guidance: 3.5
Sampler: Euler
LoRA strength: 1.70
unet_name: Flux1-dev.sft
weight_dtype: fp8_e4m3fn
clip_name1: t5xxl_fp16.safetensors
clip_name2: clip_l.safetensors
@pigletti I couldn't get any results before this. Just adding "taken at a party with flash" to my prompt suddenly made it all work, had much more effect than olympusd450 :)
@Kriima I think you have a point here. from my tests, even without the trigger word, this lora delivers the quality it promises with the right prompt. i think flux doesn't like a trigger word that it doesn't know at all for a style lora or the lora should be trained way longer
@Kriima thats interesting, I didn't have any party photos in the training data. Here's some info how I did the annotations. Maybe it helps to understand what prompts could work well, and why ChatGPT-enhanced prompts seem to do well: https://www.reddit.com/r/StableDiffusion/comments/1fa0aqv/comment/llrtjqn/
It's my first ever Style LoRA with annotated training data, so I'm quite new to this and there is certainly more room for experimentation.
@peterkickasspeter The annotation of the training data was done with ChatGPT, maybe that explains why your ChatGPT prompts look so great!
@pigletti i agree. long prompts is the way to go with flux imo
@Kriima Did you try that without the lora? I think it would look the same.
What was the lr while training this model?
lr: 1.0e-04
This bringed back some very good memories.. life was better :(
Cyclops mascot
So real 90's
scary shit how accurate this is
absolutely crazy
Use this in conjunction with the this lora.
Trust me
Dirty Cellphone - v1.0 | Stable Diffusion LoRA | Civitai
xd
Very well done, love it
Looks good. Only problem is that I have autism!!
Details
Files
Available On (3 platforms)
Same model published on other platforms. May have additional downloads or version variants.