Datasets:
File size: 10,581 Bytes
79067a6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 | {
"paper_id": "nfig",
"paper_title": "NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering",
"D1": [
{
"id": "nfig-D1-001",
"claim": "ImageNet ILSVRC 2012: 1.2M train / 50k val / 100k test images, 1000 categories, 256x256x3 resolution",
"source": "Section 4.1"
},
{
"id": "nfig-D1-002",
"claim": "FR-VAE tokenizer: VQGAN framework, DINO discriminator, DINOv2-base encoder, XQGAN implementation strategy, codebook K=4096, n=10 frequency bands, scale_factors [1,2,3,4,5,6,8,10,13,16], 680 total tokens",
"source": "Section 3.1, Section 4.1"
},
{
"id": "nfig-D1-003",
"claim": "NFIG Transformer: VAR backbone (decoder-only), depth 16, block-wise causal attention, Next-Frequency Prediction (coarse-to-fine), model sizes 310M/600M, token sequence length 680 (vs AR baseline 256)",
"source": "Section 3.2, Section 4.1"
},
{
"id": "nfig-D1-004",
"claim": "Frequency band division: n=10 frequency bands with increasing scale sequence, lower-frequency bands get fewer tokens (smaller h_i*w_i), higher-frequency bands get more tokens (larger h_i*w_i), bandwidth per band proportional to its token count ratio h_i*w_i / sum_j(h_j*w_j), with h_n=H', w_n=W'",
"source": "Section 3.2, Eq. 9"
},
{
"id": "nfig-D1-005",
"claim": "NFIG training: PyTorch, NVIDIA H100, Adam optimizer, lr=8e-5, batch_size=768, 350 epochs (VAR baseline: 200ep for 310M, 250ep for 600M, 300ep for 1B, 350ep for 2B); scaling study at 55 epochs; epoch comparison at 200/250/300/350 epochs",
"source": "Section 4.1, Section 4.2"
},
{
"id": "nfig-D1-006",
"claim": "FR-VAE loss weights: reconstruction_loss=1.0, frequency_quantized_loss=1.0, LPIPS perceptual loss=1.0, GAN loss=0.5",
"source": "Appendix B.1, Eq. 10"
},
{
"id": "nfig-D1-007",
"claim": "NFIG inference: CFG=4.5, top_k=990, 10 inference steps",
"source": "Section 4.1"
}
],
"D2": [
{
"id": "nfig-D2-001",
"claim": "f_hat_i = F^{-1}(F(f) .* M_i) — Frequency-guided Decomposer: decomposes encoder feature map f into n frequency components via FFT-based masking and inverse FFT, each mask M_i selects a specific frequency band",
"source": "Section 3.1.1, Eq. 1"
},
{
"id": "nfig-D2-002",
"claim": "f_tilde = sum_{i=1}^{n} T(f_hat_i, H', W') — Frequency-guided Composer: merges frequency components of potentially different resolutions back into a unified feature map by interpolating each to H'xW' and summing",
"source": "Section 3.1.1, Eq. 2"
},
{
"id": "nfig-D2-003",
"claim": "v_0 = argmin ||f_hat_0 - Z(v_0)||^2; R_0 = f_hat_0 - Z(v_0); for i>=1: v_i = argmin ||(R_{i-1}+f_hat_i) - Z(v_i)||^2; R_i = R_{i-1} + (f_hat_i - Z(v_i)) — Residual Token Extraction: progressive frequency-band quantization with cumulative residual R_i tracking unencoded signal through level i",
"source": "Section 3.1.2, Eq. 4"
},
{
"id": "nfig-D2-004",
"claim": "t^{(j,k)} = lookup(Z, argmin_{z in Z} ||z - v_i^{(j,k)}||_2) — Vector Quantization: converts continuous feature vectors to discrete codebook indices via nearest-neighbor L2 lookup in learnable codebook Z (KxC)",
"source": "Section 3.1.2, Eq. 5"
},
{
"id": "nfig-D2-005",
"claim": "p(T_1,...,T_n) = prod_{i=1}^{n} p(T_i | T_1,...,T_{i-1}) — Next-Frequency Autoregressive Prediction: factorizes joint token distribution as product of conditionals from low to high frequency (coarse-to-fine), with block-wise causal attention enforcing band-level autoregressive ordering",
"source": "Section 3.2, Eq. 7"
},
{
"id": "nfig-D2-006",
"claim": "sigma_i = sigma_{i-1} + (h_i*w_i / sum_j(h_j*w_j)) * sigma_max — Frequency Band Division Strategy: allocates bandwidth proportionally to token count (h_i*w_i) of each band; low-frequency bands get narrower bandwidth, high-frequency bands get wider bandwidth",
"source": "Section 3.2, Eq. 9"
},
{
"id": "nfig-D2-007",
"claim": "L = ||I - I_hat||_2^2 + ||f - f_hat||_2^2 + L_p(I) + 0.5 * L_g(I) — FR-VAE Total Loss: pixel MSE + feature MSE + LPIPS perceptual loss + 0.5x GAN adversarial loss (DINO discriminator); additional VQ codebook/commitment losses not shown",
"source": "Appendix B.1, Eq. 10"
},
{
"id": "nfig-D2-008",
"claim": "L(T, T_tilde) = -sum_{i=1}^{n} t_i * log(t_tilde_i) — NFIG Transformer Cross-Entropy Loss: standard autoregressive next-token prediction loss over all frequency-band tokens with block-wise causal masking",
"source": "Appendix B.1, Eq. 11"
},
{
"id": "nfig-D2-009",
"claim": "Block-wise Causal Attention (referenced from VAR [19]): decoder-only transformer enforces autoregressive ordering across frequency bands — tokens in lower-frequency bands (1..i-1) are visible to band i, while higher-frequency bands (i+1..n) are masked; attention within each frequency block follows VAR's block-wise scheme",
"source": "Section 3.2"
}
],
"D3": [
{
"id": "nfig-D3-001",
"claim": "Main Generation Experiment: class-conditional ImageNet 256x256 generation, train FR-VAE tokenizer then NFIG-310M Transformer (350 epochs, Adam lr=8e-5, batch=768, H100), inference with CFG=4.5 top_k=990 in 10 steps; compared against GAN (BigGAN, GigaGAN, StyleGAN-XL), Diffusion (ADM, CDM, LDM-4-G, DiT-L/2, DiT-XL/2, L-DiT-3B/7B), Mask Diffusion (MaskGIT, RCG), AR (VQVAE-2, VQGAN, ViTVQ, RQTransformer, FAR-B/H, XQGAN-310M, VAR-d16/d20); metrics: gFID, rFID, IS, Precision, Recall, params, steps, relative inference time",
"source": "Section 4.1, Section 4.2, Table 2"
},
{
"id": "nfig-D3-002",
"claim": "Ablation Study: incremental component contribution on ImageNet val 50k — baseline AR (seq_len=256) → +FR-Quantizer (seq_len=680) → +DINO-Disc → +AdaLN Transformer → +Top_k → +CFG=full NFIG; each variant trained from scratch with same FR-VAE; evaluated by rFID and gFID",
"source": "Section 4.3, Table 5"
},
{
"id": "nfig-D3-003",
"claim": "Scaling Study: NFIG-310M vs NFIG-600M trained for 55 epochs (limited compute budget), same FR-VAE tokenizer, Adam lr=8e-5 batch=768; compare FID/IS/Precision/Recall between scales to validate scaling behavior",
"source": "Section 4.2, Table 3"
},
{
"id": "nfig-D3-004",
"claim": "Epoch Efficiency Comparison: NFIG-310M trained at 200/250/300/350 epochs vs VAR-d16 (310M, 200ep) and VAR-d20 (600M, 250ep); validate that NFIG-310M at 200ep outperforms VAR-d16 and at 350ep matches VAR-d20 despite 2x fewer params; metrics: FID, IS, Precision, Recall",
"source": "Section 4.2, Table 4"
},
{
"id": "nfig-D3-005",
"claim": "Frequency Distribution Analysis: generate images with NFIG and visualize intermediate results at steps 2/4/6/8/10; compute FFT frequency spectrum at each step to show progressive low-to-high refinement; compare VQ loss across scale factors between NFIG and VAR to demonstrate more balanced feature learning",
"source": "Section 4.4, Figure 4, Figure 5"
},
{
"id": "nfig-D3-006",
"claim": "Frequency Keep Ability Analysis: compare NFIG vs VAR-16 on ImageNet using PSD (Power Spectral Density, lower is better) and FKS (Frequency Keep Score, weighted: Low 0.57, Mid 0.28, High 0.15, higher is better) across low/mid/high frequency bands",
"source": "Appendix B.2"
},
{
"id": "nfig-D3-007",
"claim": "Cross-Dataset FR-VAE Reconstruction: evaluate zero-shot reconstruction on DTD, QRCODE, Diagrams, Chest-X, CelebA-HQ, COCO, LSUN-Bedroom using ImageNet-pretrained FR-VAE; compute rFID between original and reconstructed images per dataset",
"source": "Appendix B.4"
}
],
"D4": [
{
"id": "nfig-D4-001",
"claim": "NFIG main pipeline (paper-specified generation workflow): (1) Train FR-VAE image tokenizer with frequency-guided residual quantization + VQGAN losses → (2) Train NFIG Transformer on FR-VAE ground-truth frequency tokens with cross-entropy loss → (3) Inference: generate tokens autoregressively from low to high frequency (10 steps), decode via FR-VAE decoder",
"source": "Section 4.1"
},
{
"id": "nfig-D4-002",
"claim": "Ablation experiment sequence (paper-specified): (1) Incrementally add components to baseline AR model (FR-Quantizer → DINO-Disc → AdaLN → Top_k → CFG) → (2) Train each variant from scratch with same FR-VAE tokenizer → (3) Evaluate rFID and gFID on ImageNet validation set",
"source": "Section 4.3"
},
{
"id": "nfig-D4-003",
"claim": "Scaling study sequence: (1) Train NFIG-310M for 55 epochs → (2) Train NFIG-600M for 55 epochs → (3) Compare FID/IS/Precision/Recall between the two scales under same limited compute budget",
"source": "Section 4.2"
},
{
"id": "nfig-D4-004",
"claim": "Epoch efficiency comparison sequence: (1) Train NFIG-310M for multiple epoch budgets (200, 250, 300, 350) → (2) Compare against VAR-d16 (310M, 200ep) and VAR-d20 (600M, 250ep) → (3) Validate NFIG-310M at 200ep already outperforms VAR-d16, and at 350ep matches VAR-d20 despite 2x fewer parameters",
"source": "Section 4.2"
},
{
"id": "nfig-D4-005",
"claim": "Frequency visualization sequence: (1) Generate images with NFIG, capture intermediates at steps 2/4/6/8/10 → (2) Compute FFT frequency spectrum at each step to visualize progressive low-to-high frequency refinement → (3) Compare VQ loss per scale factor between NFIG and VAR to demonstrate balanced feature learning",
"source": "Section 4.4"
},
{
"id": "nfig-D4-006",
"claim": "Frequency Keep Ability analysis sequence: (1) Generate images from both NFIG and VAR-16 → (2) Compute Power Spectral Density (PSD) for both → (3) Compute Frequency Keep Score (FKS) with weighted per-band similarity (Low 0.57, Mid 0.28, High 0.15) → (4) Compare per-band fidelity between models",
"source": "Appendix B.2"
},
{
"id": "nfig-D4-007",
"claim": "Cross-dataset FR-VAE evaluation sequence: (1) Load ImageNet-pretrained FR-VAE tokenizer → (2) Encode and decode images from each target dataset (DTD, QRCODE, Diagrams, Chest-X, CelebA-HQ, COCO, LSUN-Bedroom) → (3) Compute rFID between original and reconstructed images per dataset",
"source": "Appendix B.4"
}
]
} |