empero-ai/Homebrew-Qwen-Image-2.1-Y2K

Y2K Digicam Snapshot LoRA for Qwen-Image 2.1

A LoRA for Qwen-Image 2.1 that renders images in the Y2K digicam snapshot aesthetic of the early 2000s: direct-flash highlights, muted sensor colour cast, soft detail. Trigger with "y2kphoto" in the prompt. Trained on 160 Wikimedia Commons digicam photographs (CC BY-SA 3.0 and public domain, credits in the data section).

This LoRA is the checkpoint saved at step 500 of 1000 training steps.

Base vs step 500

Same prompt and seed, left: base Qwen-Image 2.1, right: this LoRA at step 500.

Prompt	Base	This LoRA (step 500)
y2kphoto, friends at the beach at sunset, direct flash
y2kphoto, rainy city street at night, neon reflections
y2kphoto, cat on a windowsill, soft indoor light

Brewed with Homebrew by Empero, starting from Qwen-Image 2.1.

License notice

Non-commercial use only. The base model is released under the Qwen Research License (non-commercial); this derivative and its outputs may not be used commercially.

Built with Qwen. Qwen-Image 2.1 is licensed under the Qwen Research License Agreement; this adapter and its outputs may only be used for non-commercial research or evaluation.

How to use

import torch
from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("empero-ai/Homebrew-Qwen-Image-2.1-Y2K")
image = pipe("y2kphoto, a cozy reading nook, warm light", num_inference_steps=40).images[0]
image.save("brew.png")

Qwen-Image 2.1 needs a recent diffusers (pip install git+https://github.com/huggingface/diffusers).

Samples

Prompt

y2kphoto a group of friends at the beach at sunset, direct flash, early-2000s digicam snapshot

Prompt

y2kphoto a rainy city street at night with neon signs and wet asphalt reflections

Prompt

y2kphoto a cat sitting on a windowsill, soft indoor light, early-2000s digicam snapshot

Training

Stage	Objective	Method	Data	Steps	Result
1	Supervised fine-tuning (SFT)	LoRA	160 images	1000	train loss 0.398

Stage 1 hyperparameters

objective: sft
method: lora
learning_rate: 0.0001
max_steps: 1000
effective_batch_size: 1
optimizer: adamw_8bit
scheduler: constant_with_warmup
warmup_ratio: 0.0
lora:
  rank: 16
  alpha: 16
  dropout: 0.0
  targets: default
image:
  resolution: 1024
  trigger_word: y2kphoto
  caption_dropout: 0.05

Trained on NVIDIA RTX PRO 5000 Blackwell.

Data

Dataset	Records	License	Notes
closestfriend/y2k-digicam	160	mixed per-image: CC BY-SA 3.0 / public domain (credits kept)	image

Training data was stored in ETF (Empero Trace Format) and rendered with the base model's own chat template.

Limitations

This model inherits the capabilities, biases and limitations of its base model and its training data. It can produce incorrect or inappropriate output; evaluate it for your use case before relying on it.

Credits

• Brewed with Homebrew by Empero — Independent AI research lab · Open by default · Built in Germany.
• Base model: Qwen/Qwen-Image-2.1 by Alibaba Qwen.
• Dataset: closestfriend/y2k-digicam

_{Card generated by Homebrew 0.1.0 on 2026-10-05.}