qwen-image-witcher3-velen-lora

LoRA adapter for Qwen/Qwen-Image-2512 trained on aerial cartographic terrain captures from The Witcher 3: Wild Hunt's Velen region. The adapter steers the base model toward stylized top-down cartography: muted earth tones, dense forests, river networks, marsh textures, and the distinctive flat-shaded hills of the Velen overworld.

Training config

• Rank: r=128, lora_alpha=128, bias=none, init=gaussian
• Target modules: attention proj (to_q/k/v, add_q/k/v_proj, to_out.0, to_add_out) + MLP/modulation (img_mlp.net.2, img_mod.1, txt_mlp.net.2, txt_mod.1)
• Schedule: 5 epochs over the full witcher3_velen dataset (~10k samples after repetitions); final checkpoint at step 2120
• Text-to-image: prompt embeddings via the Qwen2.5-VL processor path (pipe._get_qwen_prompt_embeds)
• Inference: DiffSynth FlowMatchScheduler (vendored) with shift=3.0, 20 steps, guidance_scale=1.0

Dataset

PNW-GM/witcher3_velen

Usage

import torch
from diffusers import QwenImagePipeline

pipe = QwenImagePipeline.from_pretrained(
    "Qwen/Qwen-Image-2512", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("PNW-GM/qwen-image-witcher3-velen-lora")

# Vendored DiffSynth FlowMatchScheduler with shift=3.0 is recommended.
# See the training repo for the exact scheduler swap.

image = pipe(
    prompt="aerial cartographic view of a marshland village in the Velen style",
    num_inference_steps=20,
    guidance_scale=1.0,
).images[0]
image.save("velen.png")

What it produces

Top-down, slightly oblique aerial views with the Velen palette: ochre and olive ground, dense pine canopies, winding rivers, swamps with reed flecks, and scattered ruined settlements. Best results when prompts describe terrain rather than characters or close-up scenes.