LuminaRS — Lightweight Recursive Art Image Generator

A novel ~90M parameter image generation model for art/illustration that runs on mobile devices (2-4GB VRAM).

Why LuminaRS?

Problem	Current Solutions	LuminaRS
Heavy models (6-12GB)	SDXL, Flux	~90M params, <500MB
Can't run mobile	Quantized SD (quality loss)	Designed small from scratch
Poor prompt adherence	SD 1.5	TRM-style recursive reasoning
No art specialization	General photo models	Art-focused training stages
Unstable training	Diffusion (score matching)	Flow matching (stable ODE)

Architecture (Novel Contributions)

1. Recursive Shared-Weight Refinement (from TRM)

Inspired by Tiny Recursive Models — beat 200x larger LLMs with 7M params.

for _ in range(T): z = z + unet(z, text, t)  # shared-weight refinement

Effective depth = T x L without Tx parameters.

2. Flow Matching (instead of Diffusion)

• v(x_t, t) = x_clean - x_noise (straight-line velocity)
• 10-12 inference steps vs 50+ for diffusion
• No score matching instability

3. ConvNeXt + MQA Cross-Attention

Depthwise 7x7 conv, Adaptive LayerNorm (time), MQA cross-attn (text), GELU MLP

4. Staged Freeze/Thaw Training

Stage	What's Trained	LR
1	All denoiser params	1e-4
2	Cross-attention only	1e-5
3	All params, joint	1e-6

VAE and CLIP always frozen.

Parameter Budget

Component	Params
Encoder	~35M
Bottleneck	~15M
Decoder	~35M
Embeds	~5M
Total trainable	~90M
VAE (frozen)	~83M
CLIP (frozen)	~303M
Inference VRAM (b=1)	~1.5-2GB

Quick Start

from luminars.model import LuminaRS
from luminars.config import LuminaRSConfig
from luminars.sampler import sample_flow
cfg = LuminaRSConfig()
model = LuminaRS(cfg)
latents = sample_flow(model, text_emb, (1,16,32,32), 12)

Files

• luminars/ -- model, config, loss, sampler, train helpers
• train.py -- main training script
• LuminaRS_Colab.ipynb -- Colab notebook

Research Foundations

• TRM (Jolicoeur-Martineau 2025): Recursive reasoning
• SnapGen (2024): Mobile UNet design
• ZigMa (2024): Mamba diffusion
• Flow Matching (Lipman 2023): Stable ODE generation
• MQA (Shazeer 2019): Multi-query attention
• ConvNeXt (Liu 2022): Modernized CNN

MIT License