🎥✨ Introducing: 90zVibez — My New Qwen LoRA (Flux Re-Train)
I’m excited to share my latest LoRA release: 90zVibez ,a re-train of my original Flux LoRA, now fully adapted and optimized for Qwen.
This project started as a personal challenge. I wanted to see if I could take the nostalgic 90s aesthetic I captured in my Flux LoRA and successfully translate it into a new checkpoint with Qwen. After a lot of experimenting, refining, and testing… I’m really happy with how the results turned out.
📼 What is 90zVibez?
90zVibez is designed to recreate the authentic look and feel of 90s candid photography—the kind of images you’d find in old shoebox photo collections, VHS freeze-frames, or 4x6 prints developed at the local drugstore.
It focuses on:
✅ Real candid emotions & natural poses
✅ Authentic 90s camera imperfections
✅ Film grain, warm tones, subtle flash artifacts
✅ Bedroom, party, street, and lifestyle nostalgia
🧠 Prompting Notes
The keyword “90zvibez” can enhance the style, but it’s not required.
You can achieve the look with normal prompting, but including the trigger word helps guide the LoRA toward that iconic analog feel faster and more consistently.
🧪 Testing Details
For these preview images, I used:
Karras sampler
Res Multi-Step
These gave the most natural film-like softness and grain while preserving character detail and lighting accuracy.
💡 Ideal Use Cases
This LoRA shines when you want:
🎉 90s party candids
📺 VHS and analog film stills
🛏️ Cozy bedroom moments
🚗 Road-trip nostalgia snapshots
📸 Disposable & point-and-shoot realism
🔥 Final Thoughts
This LoRA is all about vibes, authenticity, and emotion. It’s not meant to look perfect — it’s meant to look real, like a memory captured by a friend, not a staged photoshoot.
I hope you enjoy using 90zVibez as much as I enjoyed creating and refining it. Can’t wait to see what people make with it!
If you try it out, please share your results, I’d love to see your 90s-style generations!
https://www.youtube.com/@webdot3
https://Patreon.com/small0
Description
FAQ
Comments (8)
Tested with this a bit this morning with Qwen Image Edit 2509. Here's what I found:
- With Lightning Lora (8 step) and this at strength 1, just using the prompt "90svibez" does nothing. You have to add to your prompt "A nostalgic 90s snapshot" or something similar as the OP uses in his reference images. Even then it washes out the images significantly and I can't replicate the "dark and moody" vibe from his samples
- This is a SFW model only, NSFW prompts do not work
- Without this lora and prompting for "A nostalgic 90s snapshot" you get similar picture quality as using this lora.
I think this is just a function of QIE2509 not working well with this. Hope that helps someone else testing with 2509.
Which sampler did you use? How many steps? I had to change a few things before I got the vibe I wanted. Simply doing "A nostalgic 90s snapshot" didn't yield results for me personally.
Lightning 8 step lora, yours, tried euler/simple, euler/beta, res2s/simple with Qwen Image Edit 2509.
@MysticalOne I used Qwen Image FP8 (not edit), I also used Karras as well.
Yes, I knew you did this with Image (not edit), but those of us that use Image Edit 2509 for all our generations always test new Qwen models to see if they work with it, because I use Edit 2509 for my multiple character generations. There would have been others asking if this works with that version, so I thought I'd test myself and report to those that use it :) If you ever do an Edit 2509 version of this, I'll be happy to test it out fully.
@MysticalOne Cool! I appreciate that a lot. I'm new to Qwen as a whole, so if you have tips/tricks for LoRa optimization let me know.
@g3n Unfortunately I don't know that world of optimzation or lora creation (still have yet to do my first character lora), but I have been using Qwen Image Edit 2509 extensively for image generation and its amazing for putting 2 characters together in the same scene. The issue is always realism and consistency of characters. I know I can get better results if I had my own character loras, I will get there someday, but from 2 images of faces close up to put them in a scene together where the faces are consistent is not good with 2509. I usually have to do like 20-40 generations before I get a single one that matches the faces closely, and even then I have to Inpaint most of the time, which sucks and is a lot of extra manual effort and not scalable. Its much easier to get consistency with a single reference character, but add a second and Qwen decides to draw new characters all the time.
@MysticalOne Good insight! I will play with it a little more. I appreciate your passion and feedback. Check out my YouTube channel as well if you haven't.