Literature
Display
← All papers

Flow to the Mode: Mode-Seeking Diffusion Autoencoders for State-of-the-Art Image Tokenization

Kyle Sargent · Kyle Hsu · Justin Johnson · Li Fei-Fei · Jiajun Wu

· ICCV 2025 ↗

Computer vision9 min text · CondensedOriginal paper ↗
Contents

Overview

The supplementary material includes all training settings, evaluation conventions, additional ablations, transfer and scaling experiments, wall-clock measurements, and four full-page reconstruction comparisons.

At a glance

FlowMo is a discrete image tokenizer with transformer encoders and a conditional rectified-flow decoder. Initial end-to-end training learns a distribution of plausible reconstructions; a second stage backpropagates perceptual loss through sampled reconstructions to favor modes resembling the source image. At two matched bit budgets on 256-pixel ImageNet reconstruction, it improves reconstruction FID, PSNR, and structural similarity over the selected baselines, though not LPIPS. Downstream generation shows mixed gains, and iterative decoding is substantially slower than a conventional tokenizer.

1 Introduction

Image tokenizers determine both reconstruction quality and the latent representation available to a later generative model. FlowMo replaces a conventional spatial latent grid and adversarial decoder training with a one-dimensional sequence of binary-quantized tokens and a conditional diffusion decoder.

The central idea is that accurately modeling every possible reconstruction is not the same as reconstructing a particular source image well. Perceptual post-training and a shifted sampling schedule deliberately concentrate the conditional distribution around useful modes. The claimed advantages apply to the tested compression budgets and metrics, rather than every measure of image fidelity or generation quality.

Original figure
Figure 1: Discrete tokenizer comparison. State-of-the-art discrete tokenizers are benchmarked by encoding and reconstructing the ImageNet-1K validation dataset at 256×256256\times 256 resolution. with performance measured in reconstruction FID (rFID), which trades off against compression rate as measured in bits per pixel (BPP). Whether trained for reconstruction at a low BPP (FlowMo-Lo) or high BPP (FlowMo-Hi), FlowMo achieves state-of-the art image tokenization performance compared with the respective baselines. Moreover, FlowMo is a transformer-based diffusion autoencoder which does not use convolutions, adversarial losses, or proxy objectives from auxiliary tokenizers.
Refer to caption Refer to caption Refer to caption   Refer to caption Refer to caption
Refer to caption Refer to caption Refer to caption   Refer to caption Refer to caption
Original image OpenMagViT-V2 FlowMo-Lo (Ours) LlamaGen-32 FlowMo-Hi (Ours)
(rFID=1.17) (rFID=0.95) (rFID=0.59) (rFID=0.56)
Figure 2: Example reconstructions. Comparison of original and reconstructed images of faces and text. OpenMagViT-V2 and FlowMo-Lo are 0.07-bits per pixel tokenizers to be compared against each other. LlamaGen-32 and FlowMo-Hi are 0.22-bits per pixel tokenizers to be compared against each other. Best viewed zoomed-in in the electronic version. More comparisons are available on our website.

2 Related Work

The method builds on diffusion autoencoders [45], multimodal diffusion transformers [16], and perceptual optimization through diffusion sampling. Unlike approaches that distill a prior tokenizer or operate in its continuous latent space, FlowMo learns its pixel encoder and decoder directly.

Concurrent DiTo studies continuous diffusion tokenization, whereas FlowMo’s main experiments use discrete codes. A separate continuous variant is constructed specifically for that comparison. The learned perceptual losses remain important even though adversarial losses and an auxiliary tokenizer are absent.

3 Method

The tokenizer is trained in two phases: joint flow-based encoder–decoder training, followed by perceptual refinement with the encoder frozen. A separate masked-token generative model tests whether the resulting discrete representation is useful beyond reconstruction. Sampling choices are part of the reported reconstruction recipe.

Refer to caption
Figure 3: FlowMo architecture. FlowMo is a diffusion autoencoder which encodes images xx to a latent c^\hat{c} which is quantized to cc. Then, the model decodes a rectified flow velocity vv conditioned on cc as well as a noise level tt and noised image xtx_{t}.

3.1 Architecture

An MMDiT-style encoder jointly processes image patches and initially zero latent tokens through separate interacting streams. Lookup-free quantization maps each latent scalar to a binary value. A larger transformer decoder predicts a rectified-flow velocity from the noisy image, quantized code, and time embedding. The models use maximal-update parameterization to transfer hyperparameters across widths. The tokenizer architecture is transformer-based; its perceptual training networks are separate components.

Equations 1–3
c^=eθ​(x,c0),\hat{c}=e_{\theta}(x,c_{0}), (1)
c=q(c^)=2⋅𝟙[c^≥0]−1.c=q(\hat{c})=2\cdot\mathds{1}[\hat{c}\geq 0]-1. (2)
v=dθ​(xt,c,t).v=d_{\theta}(x_{t},c,t). (3)

3.2 Stage 1A: Mode-matching pre-training

The first stage jointly trains encoder and decoder with velocity prediction across noise levels, a perceptual loss on the one-step denoised estimate, and lookup-free quantization entropy and commitment penalties. The perceptual network is LPIPS-VGG. Noise times follow a logit-normal mixture with ten percent uniform sampling, providing more coverage near the endpoints than a plain logit-normal schedule.

Equations 4–9
xt=t​z+(1−t)​xx_{t}=tz+(1-t)x (4)
ℒflow=𝔼⁡[‖x−z−dθ​(xt,q⁡(eθ​(x)),t)‖22].\mathcal{L}_{\text{flow}}=\mathbb{E}\bigg[\big\|x-z-d_{\theta}(x_{t},q(e_{\theta}(x)),t)\big\|_{2}^{2}\bigg]. (5)
ℒperc=𝔼⁡[dperc​(x,xt+t​dθ​(xt,q⁡(eθ​(x)),t))].\mathcal{L}_{\text{perc}}=\mathbb{E}\bigg[d_{\text{perc}}(x,x_{t}+td_{\theta}(x_{t},q(e_{\theta}(x)),t))\bigg]. (6)
ℒent=𝔼[H(q(c^)]−H(𝔼[q(c^)]),\mathcal{L}_{\text{ent}}=\mathbb{E}\big[H(q(\hat{c})\big]-H(\mathbb{E}\big[q(\hat{c})\big]), (7)
ℒcommit=𝔼⁡[‖c^−q⁡(c^)‖22].\mathcal{L}_{\text{commit}}=\mathbb{E}\bigg[\big\|\hat{c}-q(\hat{c})\big\|_{2}^{2}\bigg]. (8)
ℒflow+λperc​ℒperc+λcommit​ℒcommit+λent​ℒent,\mathcal{L}_{\text{flow}}+\lambda_{\text{perc}}\mathcal{L}_{\text{perc}}+\lambda_{\text{commit}}\mathcal{L}_{\text{commit}}+\lambda_{\text{ent}}\mathcal{L}_{\text{ent}}, (9)

3.3 Stage 1B: Mode-seeking post-training

Freeze the encoder and optimize the decoder using both the flow objective and a perceptual loss on a complete generated reconstruction. Gradients pass through an eight-step sampling chain with randomly chosen times and checkpointing. A ResNet perceptual network replaces LPIPS-VGG for this stage, with sample-loss weight 0.01.

Small weights provide too little perceptual benefit; large weights can overfit the perceptual metric or destabilize training. Roughly one epoch and early stopping limit this behavior. Equation 12 and the method describe joint flow-plus-sample training, despite a later analysis sentence describing only the sample loss; the edition follows the explicit objective.

Refer to caption
Figure 4: Stage 1A. The encoder and decoder are trained end-to-end with output losses ℒperc,ℒflow\mathcal{L}_{\text{perc}},\mathcal{L}_{\text{flow}} and latent losses ℒcommit,ℒent\mathcal{L}_{\text{commit}},\mathcal{L}_{\text{ent}}.
Equations 10–12
dti​(xt)=xt+(ti+1−ti)​dθ​(xt,c,ti).d_{t_{i}}(x_{t})=x_{t}+(t_{i+1}-t_{i})d_{\theta}(x_{t},c,t_{i}). (10)
ℒsample=𝔼[dperc(x,dtn∘dtn−1∘⋯∘dt1(z))],\mathcal{L}_{\text{sample}}=\mathbb{E}\bigg[d_{\text{perc}}\left(x,d_{t_{n}}\circ d_{t_{n-1}}\circ\cdots\circ d_{t_{1}}(z)\right)\bigg], (11)
ℒflow+λsample​ℒsample.\mathcal{L}_{\text{flow}}+\lambda_{\text{sample}}\mathcal{L}_{\text{sample}}. (12)
Refer to caption
Figure 5: Stage 1B. The frozen encoder eθe_{\theta} encodes the input image to cc to condition the decoder dθd_{\theta}, which is trained via backpropagation through the entire sampling chain. We also co-train with ℒflow\mathcal{L}_{\mathrm{flow}}, which is the same as in Figure 4.

3.4 Sampling

At inference, reconstruction starts from Gaussian noise conditioned on the quantized code. A power-law timestep schedule with shift four concentrates steps at low noise while taking larger early steps. Together with limited-interval classifier-free guidance, this changes the reconstruction tradeoff. The default evaluation uses twenty-five decoding steps, so a token sequence does not decode in one forward pass.

Equations 13–14
d​xt=v⁡(xt,t)=dθ​(xt,c,t),dx_{t}=v(x_{t},t)=d_{\theta}(x_{t},c,t), (13)
(t1,…,tn)=((nn)ρ,(n−1n)ρ,…,(1n)ρ),(t_{1},...,t_{n})=\left(\left(\frac{n}{n}\right)^{\rho},\left(\frac{n-1}{n}\right)^{\rho},\dots,\left(\frac{1}{n}\right)^{\rho}\right), (14)

3.5 Stage 2: Latent generative modeling

A MaskGiT [8] model learns the discrete code distribution. This second stage evaluates whether the tokenizer supports good generation under matched model-training resources, separately from how faithfully it reconstructs known input images.

4 Experiments

The experiments evaluate reconstruction at matched nominal bits per pixel, downstream generation, a continuous-tokenizer comparison, and architecture/training ablations. Reconstruction and generation FID use different reference distributions and should not be interchanged.

4.1 Main results

On 50,000 ImageNet validation images at 256×256, FlowMo-Lo uses about 0.0703 bits per pixel and improves reconstruction FID from OpenMagViT-V2’s 1.17 to 0.95, PSNR from 21.63 to 22.07, and SSIM from 0.640 to 0.649. LPIPS is slightly worse, 0.113 versus 0.111. At about 0.219 bits per pixel, FlowMo-Hi improves LlamaGen-32’s reconstruction FID from 0.59 to 0.56 and PSNR from 24.44 to 24.93, while LPIPS worsens from 0.064 to 0.073. The bit rate counts code capacity, not a demonstrated entropy-coded file size with overhead.

With matched MaskGiT training, FlowMo-Lo yields generation FID 4.30 versus 3.73 for OpenMagViT-V2, precision 0.86 versus 0.80, and recall 0.47 versus 0.51. Inception Score and spatial FID improve, but the overall generation result is mixed. Better reconstruction does not automatically make the latent distribution easier to model.

The separate continuous variant matches DiTo’s latent size and reports reconstruction FID 0.65 versus 0.78 and PSNR 26.61 versus 24.10. Main discrete models use eight H100s, 130 or eighty pretraining epochs, and approximately one post-training epoch. Smaller exploratory configurations supply the ablations. Model width, patch size, training length, and tokenizer type must be distinguished across these comparisons.

Tokenizer FID ↓\downarrow IS ↑\uparrow sFID ↓\downarrow Prec. ↑\uparrow Rec. ↑\uparrow
OpenMagViT-V2 3.73 241 10.66 0.80 0.51
FlowMo-Lo (ours) 4.30 274 10.31 0.86 0.47
Table 2: Generation results. We compare two MaskGiT transformers trained atop two tokenizers at the same BPP.
Fully generated images - using OpenMagViT-V2 tokenizer (CFG=10.0)
Refer to caption
Refer to caption
Fully generated images - using FlowMo-Lo tokenizer (CFG=10.0)
Refer to caption
Refer to caption
Figure 6: Generated images. Example generated images from MaskGiT trained with different tokenizers. FlowMo can be used to train high-quality second-stage generative models. The corresponding class indices are identical for ease of comparison.
Model rFID ↓\downarrow PSNR ↑\uparrow SSIM ↑\uparrow LPIPS ↓\downarrow
DiTo† [10] 0.78 24.10 0.706 0.102
FlowMo (ours) 0.65 26.61 0.791 0.054
Table 3: DiTo comparison. Comparison with concurrent work DiTo [10]. †Results from the original paper.
BPP Model Tokens per image Vocab size rFID↓\downarrow PSNR↑\uparrow SSIM↑\uparrow LPIPS↓\downarrow
0.006 TiTok-L-32 [61] 32 2122^{12} 2.21 15.60 0.359 0.322
0.012 TiTok-B-64 [61] 64 2122^{12} 1.70 16.80 0.407 0.252
0.023 TiTok-S-128 [61] 128 2122^{12} 1.71 17.52 0.437 0.210
0.055 LlamaGen-16 [51] 256 2142^{14} 2.19 20.67 0.589 0.132
MaskBiT† [55] 256 2142^{14} 1.37 21.5 0.56 -
0.062 Cosmos DI-16x16 [61] 256 ≈216\approx 2^{16} 4.40 19.98 0.536 0.153
0.070 OpenMagViT-V2 [38] 256 2182^{18} 1.17 21.63 0.640 0.111
FlowMo-Lo (ours) 256 2182^{18} 0.95 22.07 0.649 0.113
0.203 ViT-VQGAN† [59] 1024 2132^{13} 1.28 - - -
0.219 LlamaGen-32 [51] 1024 2142^{14} 0.59 24.44 0.768 0.064
FlowMo-Hi (ours) 1024 2142^{14} 0.56 24.93 0.785 0.073
0.249 Cosmos DI-8x8 [61] 1024 ≈216\approx 2^{16} 0.87 24.82 0.763 0.072
Table 1: Tokenization results. Horizontal lines separate tokenizers trained at different BPPs. FlowMo achieves state-of-the-art performance at multiple BPPs compared to existing state-of-the-art tokenizers. †Results from the original paper..

4.2 Analysis

The smaller configuration’s ablations support fine image patches, end-to-end flow training, perceptual supervision, endpoint-aware time sampling, shifted decoding, and guidance. Removing perceptual supervision raises reconstruction FID from 2.87 to 13.86 even while PSNR improves, exposing the difference between pixel distortion and perceptual distribution quality. Finite scalar quantization also improves some pairwise metrics but worsens reconstruction FID relative to lookup-free quantization.

Post-training improves FlowMo-Lo from 1.10 to 0.95 reconstruction FID and FlowMo-Hi from 0.73 to 0.56, with PSNR and LPIPS also improving within each model. Repeated reconstructions remain stochastic, with more variance in backgrounds in the illustrated example. Tokenizer guidance drops the latent conditioning during training; it does not use class labels in the tokenizer.

Refer to caption Refer to caption
Original image Reconstructed (discolored)
Figure 7: Noise schedule ablation. Training FlowMo with a logit-normal noise schedule [16] results in discoloration.
Model rFID↓\downarrow PSNR↑\uparrow LPIPS↓\downarrow
FlowMo (fewer params.) 2.87 20.71 0.15
with doubled patch size 6.39 19.94 0.17
with MSE-trained encoder 3.82 21.40 0.15
without perceptual loss 13.86 22.11 0.21
with FSQ quantization 3.14 21.31 0.14
with logit-normal schedule 4.08 16.45 0.21
without shifted sampler 3.42 20.25 0.16
without guidance 3.28 20.67 0.16
Table 4: Stage 1A Ablation. Deviating from FlowMo design choices compromises either PSNR or rFID. We prioritize rFID in our model due to its correlation with perceptual quality.
Model rFID↓\downarrow PSNR↑\uparrow LPIPS↓\downarrow
FlowMo-Lo 1.10 21.38 0.134
FlowMo-Lo (post-trained) 0.95 22.07 0.113
FlowMo-Hi 0.73 24.02 0.086
FlowMo-Hi (post-trained) 0.56 24.93 0.073
Table 5: Stage 1B ablation. Without the post-training stage, performance is considerably worse.
Refer to caption Refer to caption Refer to caption
Original image Reconstructed Variance heatmap
Figure 8: Multimodal reconstruction. After post-training, FlowMo reconstruction remains multimodal, but biased towards preserving the perceptually relevant details of the image, which manifests here by the variance concentrating in the background.

Limitations.

The main limitation is iterative inference: twenty-five decoding steps require many model evaluations, and backpropagation through sampling also makes post-training expensive. The model may overfit its perceptual objective, prompting early stopping. Later timing measurements quantify the decoding cost; proposed distillation speedups are not demonstrated here.

5 Conclusion

FlowMo shows that a directly trained transformer diffusion autoencoder can be competitive for discrete perceptual image reconstruction at the tested bit budgets. Its strongest evidence is reconstruction quality; LPIPS, downstream generation, and decoding speed expose tradeoffs. The work acknowledges ONR MURI support.

Appendix A Training setup and hyperparameters

The supplementary setup records optimizer settings, losses, architecture sizes, token factorization, and separate recipes for tokenizer pretraining, post-training, and latent generation.

A.1 Stage 1A - Tokenizer pre-training

Lookup-free quantization uses relatively low commitment weight, and token factorization makes the entropy calculation manageable. Hyperparameters are selected on a smaller model and transferred to the larger configurations. Table 6 lists a 517-million-parameter exploratory model and 945-million-parameter main models, with four-pixel patches for the latter.

The high-rate representation is listed as 256 tokens of 56 bits in the architecture table and 1,024 tokens of fourteen bits in the benchmark table; both total 14,336 bits, consistent with code factorization. The main prose says only decoder width increases, whereas Table 6 gives a single hidden-width value, so the table does not independently resolve encoder width. All original settings are retained.

Hyperparameter FlowMo (fewer params) FlowMo-Lo FlowMo-Hi
Learning rate 0.0001 - -
Batch size 128 - -
Weight decay 0 - -
Num. epochs 40 130 80
λe​n​t\lambda_{ent} 0.0025 - -
λc​o​m​m​i​t\lambda_{commit} 0.000625 - -
λl​p​i​p​s\lambda_{lpips} 0.1 - -
Hidden size (μ​P\mu P width) 768 1152 1152
MLP ratio 4 - -
Encoder patch size 8 4 4
Decoder patch size 8 4 4
Encoder depth 8 - -
Decoder depth 16 - -
Latent sequence length 256 - -
Latent token size 18 18 56
Codebook size for entropy loss 9 9 14
Total number of parameters (×106)(\times 10^{6}) 517 945 945
Table 6: Training hyperparameters for tokenizer training. ‘-’ means hyperparameters are identical between FlowMo Small, A, and B configurations

A.2 Stage 1B - Tokenizer post-training

Replacing the one-step perceptual loss with a ResNet network does not match full-chain post-training in the loss-weight sweep: reconstruction FID is 1.28 for the sampled-chain loss versus a best listed 1.60 for the one-step alternatives. Post-training uses batch size 64, learning rate 0.00005, gradient accumulation and checkpointing, and about one epoch. The retained timestep equation normalizes cumulative random draws to set the sampled integration times.

Loss type Loss weight rFID
ℒsample\mathcal{L}_{\textrm{sample}} 0.01 1.28
ℒperc\mathcal{L}_{\textrm{perc}} 0.05 2.57
ℒperc\mathcal{L}_{\textrm{perc}} 0.01 2.38
ℒperc\mathcal{L}_{\textrm{perc}} 0.005 2.16
ℒperc\mathcal{L}_{\textrm{perc}} 0.001 1.67
ℒperc\mathcal{L}_{\textrm{perc}} 0.0005 1.61
ℒperc\mathcal{L}_{\textrm{perc}} 0.0001 1.81
ℒperc\mathcal{L}_{\textrm{perc}} 0.00025 1.64
ℒperc\mathcal{L}_{\textrm{perc}} 0.0005 1.61
ℒperc\mathcal{L}_{\textrm{perc}} 0.00075 1.60
Table 7: ℒsample\mathcal{L}_{\textrm{sample}} ablation. We show the result of training with ℒsample\mathcal{L}_{\textrm{sample}} in the first row. We then compare with ℒperc\mathcal{L}_{\textrm{perc}} at various weights. We conduct two sweeps, first to find the optimal loss weight order of magnitude (middle rows) and then to fine-tune the loss weight (bottom rows). Regardless, it cannot match the performance of ℒsample\mathcal{L}_{\textrm{sample}} Backpropagation through a complete sampling chain is important.
Equation
Original supplementary sampling-time equation, reproduced from PDF page 12.

A.3 Stage 2 - Generative model training

The second-stage transformer has hidden size 1,024, twenty-eight layers, and 397 million parameters. Both tokenizers supply 256 eighteen-bit codes, split into 512 nine-bit tokens for masked-token training. Models train for 300 epochs at batch size 128 and learning rate 0.0001. Sampling uses 64 steps and separately tuned temperatures. The stated generative guidance is 1.5, while Figure 6’s qualitative grid labels CFG 10.0; the figure should not be assumed to show the exact quantitative evaluation setting.

Appendix B Evaluation setup and hyperparameters

Tokenizer evaluation uses guidance weight 1.5 within rectified-flow noise levels approximately 0.145–0.505. Reconstruction metrics compare the full 50,000-image validation set with its reconstructions. Generation uses 50,000 samples and training-set reference statistics. These conventions differ from the five-thousand-sample tradeoff experiment and from the perceptual post-training loss itself.

Appendix C Additional Experiments

Additional experiments examine perceptual tradeoffs, new datasets and resolutions, scaling, precision and recall, other model variants, and measured encoding/decoding times.

C.1 Rate-distortion-perception tradeoff

Changing the norm of the initial Gaussian noise traces a tradeoff between reconstruction FID and PSNR, rather than reproducing the simultaneous gains of the proposed recipe. Increasing apparent source-noise likelihood is not the same as maximizing full image likelihood, which also depends on the flow Jacobian. Figure 9 uses 5,000 samples per setting; the experiment does not evade general rate–distortion–perception tradeoffs.

Original figure
Figure 9: Analysis of FlowMo with respect to the rate-distortion-perception tradeoff [4]. Varying x1x_{1} likelihood sweeps out different rFID/PSNR curves at different rates. For this study, we use 5,000 samples for all metrics to make it computationally feasible.

C.2 Dataset and resolution generalization.

Patchwise diffusion and blending extend evaluation from 256 to 512 resolution and to OpenImages and Food-101. Table 8 labels the evaluated model FlowMo-Hi, despite the surrounding prose naming FlowMo-Lo. It improves reconstruction FID over Cosmos DI-8×8 on five of six listed dataset/resolution combinations, but is worse on Food-101 at 256 resolution, 1.39 versus 1.05. These are limited transfer tests rather than universal resolution-independent superiority.

ImageNet-1K OpenImages Food-101
Resolution 256 512 256 512 256 512
Cosmos DI-8x8 0.87 0.39 0.85 0.52 1.05 0.94
FlowMo-Hi 0.55 0.30 0.69 0.40 1.39 0.68
Table 8: Multiple resolutions. FlowMo can perform encoding and decoding at arbitrary resolutions following a patchwise diffusion strategy known in prior work [25].

C.3 Tokenizer scaling.

At a fixed 200,000 training steps, increasing width raises model size from 260 to 710 million parameters and improves reconstruction FID from 7.77 to 3.84, PSNR from 20.84 to 21.70, and LPIPS from 0.169 to 0.152. The comparison holds steps fixed, so larger models also use more computation; it is not a fixed-FLOP scaling result.

μ​P\mu P width # Params (×106)\times 10^{6}) rFID ↓\downarrow PSNR ↑\uparrow LPIPS ↓\downarrow
2 260 7.77 20.84 0.169
3 367 5.31 21.28 0.160
4 517 4.45 21.60 0.155
5 710 3.84 21.70 0.152

C.4 “Mode-seeking,” precision and recall

When reconstructions and references correspond image by image, post-training improves both measured precision and recall. Against separate training-set statistics, precision rises from 0.734 to 0.766 while recall falls from 0.660 to 0.634. This supports the interpretation that the decoder favors higher-fidelity modes while reducing some distributional diversity, rather than proving diversity is always preserved.

vs. train stats. vs. val stats.
Model name Prec.↑\uparrow Rec.↑\uparrow Prec.↑\uparrow Rec.↑\uparrow
FlowMo (fewer params) (no posttrain) 0.734 0.660 0.974 0.988
FlowMo (fewer params) 0.766 0.634 0.993 0.991

C.5 More results with post-training

Additional post-training examples improve the smaller discrete model from 2.02 to 1.21 reconstruction FID and the continuous variant from 0.74 to 0.65, with accompanying PSNR and LPIPS gains. These configuration-specific numbers supplement the main models and should not be substituted for the earlier differently trained ablation baseline.

Model name rFID ↓\downarrow PSNR ↑\uparrow LPIPS ↓\downarrow
FlowMo (fewer params) (no post-train) 2.02 21.32 0.143
FlowMo (fewer params) 1.21 22.20 0.120
FlowMo-Continuous (no post-train) 0.74 25.65 0.064
FlowMo-Continuous 0.65 26.61 0.054

C.6 Wall clock times

The timing table reports 0.050 seconds to encode and 2.391 seconds to decode an image for the model labeled FlowMo-B, versus 0.021 and 0.038 seconds for LlamaGen-32. The listed small model takes 0.011 and 0.322 seconds. The full diffusion decoder is therefore about 63 times slower in that table, while encoding is much closer. The local table does not identify a complete timing hardware/batch protocol; the suggested five-to-tenfold distillation speedup is future work.

Model name Encode 1 image (s) Decode 1 image (s)
LlamaGen-32 0.021 0.038
FlowMo-B 0.050 2.391
FlowMo-B (small) 0.011 0.322

Appendix D Extended visualization

Four full-page grids compare original images with low-rate OpenMagViT-V2/FlowMo and high-rate LlamaGen/FlowMo reconstructions. Every image and column is retained at original resolution, including text, faces, animals, transparent objects, and fine textures. The authors state these examples were not cherry-picked; they illustrate local differences alongside the aggregate metrics.

Refer to caption Refer to caption Refer to caption
Original image Reconstructed Reconstructed
(OpenMagViT-V2 [38]) (FlowMo-A)
Figure 10: Tokenizer comparison.
Refer to caption Refer to caption Refer to caption
Original image Reconstructed Reconstructed
(OpenMagViT-V2 [38]) (FlowMo-A)
Figure 11: Tokenizer comparison, continued.
Refer to caption Refer to caption Refer to caption
Original image Reconstructed Reconstructed
(LlamaGen-32 [51]) (FlowMo-B)
Figure 12: Tokenizer comparison, continued.
Refer to caption Refer to caption Refer to caption
Original image Reconstructed Reconstructed
(LlamaGen-32 [51]) (FlowMo-B)
Figure 13: Tokenizer comparison, continued.
Concept demonstration

Open image ↗