Update README.md

README.md

@@ -17,11 +17,11 @@ datasets:
 
 # StyleVAR: Controllable Image Style Transfer via Visual Autoregressive Modeling
 
-Reference-based image style transfer built on Visual Autoregressive Modeling (VAR). Given a content image and a style image, StyleVAR generates a stylized output autoregressively over 10 scales ($1{\times}1 \to 16{\times}16$, $L=680$ tokens) using a **Blended Cross-Attention** mechanism: style and content features act as queries over the target's own autoregressive history, preserving content structure while absorbing style texture.
+Reference-based image style transfer built on Visual Autoregressive Modeling (VAR). Given a content image and a style image, StyleVAR generates a stylized output autoregressively over 10 scales (1×1 → 16×16, L=680 tokens) using a **Blended Cross-Attention** mechanism: style and content features act as queries over the target's own autoregressive history, preserving content structure while absorbing style texture.
 
 The model is trained in two stages from a pretrained vanilla VAR checkpoint:
 - **Stage 1 — SFT.** Supervised fine-tuning on 267,710 paired (content, style, target) triplets.
-- **Stage 2 — GRPO.** Reinforcement fine-tuning with Group Relative Policy Optimization against a DreamSim-based perceptual reward, using LoRA adapters ($r{=}256$) and Per-Action Normalization Weighting (PANW) to rebalance credit across the $256\times$ token imbalance between coarse and fine scales.
+- **Stage 2 — GRPO.** Reinforcement fine-tuning with Group Relative Policy Optimization against a DreamSim-based perceptual reward, using LoRA adapters (rank r=256) and Per-Action Normalization Weighting (PANW) to rebalance credit across the 256× token imbalance between coarse and fine scales.
 
 - **Code:** https://github.com/Senfier-LiqiJing/StyleVAR
 - **Paper (OpenReview):** https://openreview.net/forum?id=UHW3PgLUsa
@@ -33,7 +33,7 @@ The model is trained in two stages from a pretrained vanilla VAR checkpoint:
 
 | File | Purpose | Notes |
 |---|---|---|
-| `vae_ch160v4096z32.pth` | Frozen multi-scale VQ-VAE tokenizer | Inherited from VAR (depth-16, $C_\text{vae}{=}32$, $V{=}4096$, $\text{ch}{=}160$). Shared between SFT and GRPO. |
+| `vae_ch160v4096z32.pth` | Frozen multi-scale VQ-VAE tokenizer | Inherited from VAR (depth-16, C_vae=32, V=4096, ch=160). Shared between SFT and GRPO. |
 | `StyleVAR_SFT.pth` | Stage 1 supervised fine-tuning checkpoint | Plain state dict — no LoRA, no optimizer state. Use this for the SFT baseline. |
 | `StyleVAR-GRPO.pth` | Stage 2 GRPO-refined checkpoint | GRPO LoRA deltas already merged into the base weights. Drop-in replacement for `StyleVAR_SFT.pth`. |
 
@@ -122,7 +122,7 @@ with torch.no_grad():
 | Backbone | VAR depth-20 transformer (~600M params) |
 | Trainable | Full transformer (VQ-VAE frozen) |
 | Epochs | 10 |
-| Learning rate | $5\times10^{-4}$ (epochs 1-6) → $1\times10^{-4}$ (epochs 7-10) |
+| Learning rate | 5×10⁻⁴ (epochs 1-6) → 1×10⁻⁴ (epochs 7-10) |
 | Batch size | 128 (with gradient accumulation) |
 | Hardware | 1× NVIDIA RTX 4090 (48GB) |
 
@@ -130,15 +130,15 @@ with torch.no_grad():
 
 | Setting | Value |
 |---|---|
-| LoRA | $r{=}256$, $\alpha/r{=}2$ on every attention ($W_Q^\text{target}$, $W_{QKV}^\text{cond}$, $W_\text{proj}$) and FFN linear — 131M trainable (18.2% of backbone) |
-| Reward | $R = -\lambda \cdot \text{DreamSim}(\hat{x}, x)$, $\lambda{=}5.0$ |
-| Group size $G$ | 16 |
-| Sampling | top-$k{=}900$, top-$p{=}0.96$ |
-| Clip / KL / PANW | $\varepsilon{=}0.2$, $\beta{=}0.1$, $\gamma{=}0.7$ |
-| Optimizer | AdamW, lr $1\times10^{-5}$, wd 0.01, $(\beta_1, \beta_2){=}(0.9, 0.95)$, FP32 |
-| Iterative merge | peak-triggered, $\tau_\text{gain}{=}0.05$ / $\tau_\text{patience}{=}50$ / 300-step cool-down |
+| LoRA | rank r=256, α/r=2 on every attention (W_Q_target, W_QKV_cond, W_proj) and FFN linear — 131M trainable (18.2% of backbone) |
+| Reward | R = −λ · DreamSim(x̂, x), λ=5.0 |
+| Group size G | 16 |
+| Sampling | top-k=900, top-p=0.96 |
+| Clip / KL / PANW | ε=0.2, β=0.1, γ=0.7 |
+| Optimizer | AdamW, lr 1×10⁻⁵, wd 0.01, (β₁, β₂)=(0.9, 0.95), FP32 |
+| Iterative merge | peak-triggered, τ_gain=0.05 / τ_patience=50 / 300-step cool-down |
 | Emergency merge | raw KL > 2.0, 50-step cool-down |
-| Hardware | 1× NVIDIA RTX 4090 (48GB), physical batch 16, $G{=}16$ serial rollouts |
+| Hardware | 1× NVIDIA RTX 4090 (48GB), physical batch 16, G=16 serial rollouts |
 
 ---
 
@@ -149,15 +149,15 @@ Both stages use a concatenation of two paired style-transfer datasets:
 - **OmniStyle-150K** — 143,992 (content, style, target) triplets.
 - **ImagePulse-StyleTransfer** — 137,886 triplets.
 
-Combined into a single **267,710-sample** training set with a 95/5 train/val split. SFT applies rotation + brightness to content and random cropping to style; GRPO rollouts are performed without augmentation so that the conditioning signal is deterministic across the $G$ samples in a group.
+Combined into a single **267,710-sample** training set with a 95/5 train/val split. SFT applies rotation + brightness to content and random cropping to style; GRPO rollouts are performed without augmentation so that the conditioning signal is deterministic across the G samples in a group.
 
 ---
 
 ## Results
 
-Evaluation on three benchmarks spanning in-, near-, and out-of-distribution regimes. Arrows: higher $\uparrow$ / lower $\downarrow$ is better. **Best** in bold, <ins>second best</ins> underlined.
+Evaluation on three benchmarks spanning in-, near-, and out-of-distribution regimes. Arrows: higher ↑ / lower ↓ is better. **Best** in bold, <ins>second best</ins> underlined.
 
-| Dataset | Method | Style Loss $\downarrow$ | Content Loss $\downarrow$ | LPIPS $\downarrow$ | SSIM $\uparrow$ | DreamSim $\downarrow$ | CLIP Sim $\uparrow$ | Infer (s) $\downarrow$ |
+| Dataset | Method | Style Loss ↓ | Content Loss ↓ | LPIPS ↓ | SSIM ↑ | DreamSim ↓ | CLIP Sim ↑ | Infer (s) ↓ |
 |---|---|---|---|---|---|---|---|---|
 | **OmniStyle**    | AdaIN (baseline) | 0.0625 | 198.3449 | 0.7506 | 0.1421 | 0.6522 | 0.6555 | **0.0079** |
 |                  | StyleVAR (SFT)   | <ins>0.0468</ins> | <ins>116.3569</ins> | <ins>0.4743</ins> | <ins>0.3975</ins> | <ins>0.2276</ins> | <ins>0.8704</ins> | 0.4031 |