OptiScene-Qwen3-8B

An open-source reproduction of OptiScene (OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization, NeurIPS 2025, arXiv:2506.07570) built on Qwen/Qwen3-8B and trained on 8 × NVIDIA A100-SXM4-80GB GPUs.

Repository Structure

Root directory (/): Full 16 GB merged bfloat16 model weights after all three training stages (Stage 1 SFT + Stage 2 DPO-Stage1 + Stage 3 DPO-Stage2), ready for direct AutoModelForCausalLM.from_pretrained("Prisoner/OptiScene-Qwen3-8B") or vLLM inference.
lora_adapters/stage1_sft/: Stage 1 SFT LoRA adapter ($r=16, \alpha=32$).
lora_adapters/stage2_dpo1/: Stage 2 DPO-Stage1 LoRA adapter ($r=32, \alpha=16$, trained on top of Stage 1 merged model).
lora_adapters/stage3_dpo2/: Stage 3 DPO-Stage2 LoRA adapter ($r=32, \alpha=16$, trained on top of Stage 2 merged model).
eval_results_summary.json: Quantitative evaluation results across 120 stratified test scenes (Bedroom, Living room, Kitchen, Bathroom).
sample_response.txt: Sample 3D living-room layout output generated by the final model (OOR = 0.0099).

Training Summary

Stage	Dataset	Samples (80/20 split)	LoRA Config	Epochs	Final Metric
Stage 1: SFT (w/ Reasoning)	`sft_prompts.json`	12,253	$r=16, \alpha=32$	1	`train_loss = 0.3144`, `token_acc = 95.54%`
Stage 2: DPO Stage 1	`dpo_prompts.json`	7,951	$r=32, \alpha=16$	10 (1,990 steps)	`train_loss = 0.2468`, `rewards/margins = 8.05`
Stage 3: DPO Stage 2	`dpo2_prompts.json`	8,410	$r=32, \alpha=16$	10 (2,110 steps)	`train_loss = 0.058`, `rewards/margins = 9.86`

Quantitative Evaluation (120 Stratified Test Scenes)

Evaluated across 30 scenes per room type (Bedroom, Living room, Kitchen, Bathroom) using the Object Overlap Rate ($\text{OOR} \downarrow$) metric defined in Appendix D.2 of the OptiScene paper:

$\text{OOR} = \frac{\sum_{i=1}^{N}\sum_{j=i+1}^{N} \text{Area}(\text{Intersection}(B_i, B_j))}{\sum_{i=1}^{N} \text{Area}(B_i)}$

Model Stage	Overall OOR ↓	Bedroom OOR ↓	Living room OOR ↓	Kitchen OOR ↓	Bathroom OOR ↓	Valid JSON % ↑
Stage 1: SFT (w/ Reasoning)	`0.1305`	`0.2028`	`0.1478`	`0.1251`	`0.0464`	100.0%
Stage 2: SFT + DPO Stage 1	`0.0762` (-41.6%)	`0.0933`	`0.0756`	`0.1146`	`0.0212`	100.0%
Stage 3: OptiScene (SFT + DPO1 + DPO2)	`0.0706` (-45.9%)	`0.0799` (-60.6%)	`0.0774`	`0.0927` (-25.9%)	`0.0325`	100.0%

Quickstart Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Prisoner/OptiScene-Qwen3-8B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))