Pith. sign in

REVIEW 3 major objections 8 minor 2 cited by

Instella-T2I: Pushing the Limits of 1D Discrete Latent Space Image Generation

T0 review · 3 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 128 binary tokens now generate competitive 1024x1024 images

desk verdict Genuinely new 128-binary-token T2I system with real efficiency wins, but 'competitive' rests on a Table 1 whose protocol mismatches are admitted in the appendix. read the letter →

arxiv 2506.21022 v1 pith:3NLVTSSA submitted 2025-06-26 cs.CV

Ze Wang , Hao Chen , Benran Hu , Jiang Liu , Ximeng Sun , Jialian Wu , Yusheng Su , Xiaodong Yu
show 2 more authors
Emad Barsoum Zicheng Liu
This is my paper
classification cs.CV
keywords text-to-imagegenerationbinarylatentspace1Dtokenizerdiffusionmodelauto-regressivetokencompressionefficienttrainingdiscrete
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a 1024×1024 image can be faithfully represented by a sequence of just 128 discrete tokens, each token a 64-bit binary vector, and that this compact 1D latent is enough to support both diffusion-based and token-by-token autoregressive text-to-image generation at quality competitive with much larger, token-hungry baselines. If true, the work collapses the token budget for high-resolution image generation by up to a factor of 32 relative to standard VQ-VAE tokenizers, and it does so while keeping training under 200 GPU-days on a single 8-GPU node and inference below half a second for the diffusion model. The significance is that discrete image tokens become cheap enough to live inside ordinary next-token language-model architectures, and one tokenizer serves multiple resolutions and both major generative paradigms. All reported models are trained from scratch on public data with no post-training refinement.

What carries the argument

The machine that carries the argument is the 1D binary image tokenizer (1D-BIT) coupled with a Bernoulli latent-space diffusion/AR pair. 1D-BIT works by feeding image patches and $k$ learnable latent tokens through joint transformer blocks, discarding the image tokens, then quantizing the latents with a sigmoid nonlinearity followed by Bernoulli sampling to produce a binary code $\tilde{z}\in\{0,1\}^{k\times\hat{c}}$; the decoder reconstructs the image from mask tokens that attend to those binary latents. Two properties do the heavy lifting: the binary vector per token multiplies representational capacity over one-hot selection, and the multi-stage training (perceptual plus L1 loss first, then multi-resolution decoding, then adversarial fine-tuning with the encoder frozen) prevents code collapse and keeps high-frequency detail. The same bitstream is then modeled either as a continuous-time Bernoulli diffusion process, where the network predicts a flip probability via $\sigma(f_\theta(z_t,t,c))\oplus z_t$, or as an autoregressive sequence with per-bit BCE loss, so one tokenizer serves both paradigms.

What would settle it

Re-run the released checkpoints through the exact official implementations of GenEval, ImageReward, and the CLIP-score pipeline with the same prompts, resizing, and sampling settings used to produce the baselines' published numbers; if any score moves by more than the typical benchmark variance (roughly ±0.01 for CLIP, ±0.05 for GenEval), the "competitive with modern models" claim is not established. Alternatively, reproduce Table 1 with the baselines' original evaluation code and compare directly.

Watch

Extended reading notes

Core claim

The central discovery is that replacing one-hot codebook tokens with Bernoulli-sampled binary vectors in a 1D latent space removes the two bottlenecks that previously forced discrete tokenizers to use thousands of tokens: the representational capacity ceiling of a single codebook entry and the rigid 2D grid redundancy. The 1D binary image tokenizer (1D-BIT) encodes an image into $k=128$ latent tokens of $\hat{c}=64$ bits, and the same latent decodes to 512, 768, or 1024 resolution depending on the number of replicated mask tokens. On this latent, the paper trains a continuous-time Bernoulli diffusion model (XOR-flip prediction, 20-step sampling) and a causal autoregressive model (per-token binary cross-entropy), both conditioned on text features from a frozen decoder-only LLM fused into every transformer block. The reported results are GenEval 0.64, CLIP score 0.332, and ImageReward 0.90 for the 1.2B diffusion model, with the 0.8B AR model close behind, against baselines that use 16–32 times more tokens and orders of magnitude more training compute.

Load-bearing premise

The load-bearing premise is that the evaluation protocol used for Table 1 yields scores directly comparable to the baselines' published numbers; the paper itself concedes in Appendix D.2 that "potential mismatches about the detailed evaluation protocols across previous works" make this uncertain, and the entire competitiveness claim depends on that comparability.

Editorial extensions

If this is right

  • One 128-token bitstream decodes natively to 512, 768, and 1024 resolutions, so a higher-resolution image costs almost the same generation time as a lower-resolution one.
  • The 1.2B diffusion model samples a 1024×1024 image in under 0.5 seconds on one GPU with 20 steps and no timestep distillation, narrowing the efficiency gap with one-step GAN methods.
  • The autoregressive model reaches GenEval 0.46 on the same latent, matching strong diffusion baselines while using 16–32× fewer tokens than VQ-VAE-style AR models and generating in seconds rather than minutes.
  • Training all reported models (tokenizer plus 1B diffusion and AR generators) fits within 200 GPU-days on a single 8-GPU node with a global batch of 4096, using only public data.
  • Because the same binary latent is naturally consumable by next-token prediction, the approach opens the door to unified image-language understanding and generation within a single LLM-style architecture.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the 128-token binary budget holds when scaled to larger models and richer data, the dominant cost of autoregressive image generation shifts from token count to text sequence length; image tokens would cost roughly 3% of a VQ-VAE budget, making joint image-text next-token prediction as cheap as pure text generation.
  • The multi-resolution decoder suggests a testable extension: train the decoder on a continuous range of RoPE-interpolated resolutions and probe where reconstruction fidelity degrades; the paper only claims the three trained resolutions, so graceful behavior at, say, 896 or 1152 is an open empirical question.
  • Because both the diffusion flip-probability and the AR per-bit probability are Bernoulli quantities over the same latent, a natural experiment is to initialize one generator from the other's weights or distill the diffusion sampler into the AR model; the close GenEval and CLIP agreement hints they may share a common learned distribution.
  • The paper's own Appendix E admits weaker FID and reconstruction artifacts in high-frequency textures; if those artifacts persist at higher resolutions, the practical ceiling of fully-transformer binary tokenizers may be reached before the token-count gains are fully realized.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. The paper introduces Instella-T2I, a text-to-image system built around a 1D binary image tokenizer (BIT). BIT encodes a 1024x1024 image into 128 binary vectors of dimension 64 and can decode that same latent to multiple resolutions. The authors train 1B-scale diffusion and autoregressive models on open data, using a frozen OLMo-1B text encoder, and report competitive scores on GenEval, CLIP, and ImageReward with 16-32x fewer tokens than standard VQ-VAE systems and training in about 200 GPU-days. The diffusion component uses a continuous-time Bernoulli forward process with BCE loss and CFG; the AR component uses causal token-by-token binary prediction. The paper also reports reconstruction results for BIT and efficiency comparisons.

Significance. If the headline results hold under a common evaluation protocol, the work is a significant systems contribution: it is, to my knowledge, the first demonstration that 128 discrete binary tokens suffice for 1024x1024 text-to-image generation in both diffusion and AR paradigms, with multi-resolution decoding, sub-second diffusion inference, and a training budget an order of magnitude below typical 1B-scale T2I pretraining. The paper is transparent in reporting data sources, prompts, sampling algorithms, and hardware, and it explicitly lists limitations (Appendix E) including reconstruction artifacts and weaker FID. The methodological novelty is modest, since the binary latent diffusion builds directly on Wang et al. (2023a) and the 1D tokenizer follows the TiTok/SoftVQ line, but the combination and scaling results are still valuable. The main weakness is that the quantitative comparisons to prior models are not performed under a common protocol, which leaves the central 'competitive' claim unproven.

major comments (3)
  1. [Table 1; Appendix D.2] The headline claim of competitive performance is not established because the baseline numbers in Table 1 were not produced under the same evaluation protocol. Appendix D.2 specifies Instella's protocol (CLIP: ViT-G/14 on 30K MSCOCO images resized to 256x256; ImageReward: 100 prompts with one image each; GenEval: 553 prompts with 4 images each) and then concedes 'potential mismatches about the detailed evaluation protocols across previous works.' The baselines are quoted from their original publications rather than re-measured under this protocol. The margins are small enough to be affected by protocol choices: Instella Diff's GenEval 0.64 is 0.05 below SD3's 0.69, and its CLIP 0.332 is within 0.003 of SDXL (0.335) and Emu3 (0.333), while ImageReward 0.900 is above the listed baselines. A common-protocol re-evaluation of SDXL, SD3, and Emu3, or a corresponding weakening of the claim, is required before the central result can be accepted.
  2. [Table 1, Instella AR row] The abstract promises competitive performance in both diffusion and auto-regressive generation, but the auto-regressive model's overall GenEval of 0.46 is below SDv2.1 (0.50), PixArt-σ (0.52), SDXL (0.55), SD3-Medium (0.69), and Emu3 (0.54) in the same table. It is only on par with SDv1.5 (0.43) and PixArt-α (0.48). The text's statement that the AR model is 'comparable to powerful diffusion models' therefore overstates the evidence; the claim should be qualified to 'competitive with small diffusion baselines' or the AR model should be evaluated under the same protocol as the baselines.
  3. [Table 3; Appendix D.2; Appendix E] The FID column labeled 'FID-30K' is inconsistent with the protocol in Appendix D.2, which generates images from 3K MSCOCO prompts and also says 'val2024' while comparing to 'val2014'. Along with the 30K/3K mismatch, the baseline FID values are not re-measured under this protocol, and Appendix E concedes that the model 'does not excel in FID scores.' As a result, the FID numbers in Table 3 cannot support the competitive-quality claim; the efficiency comparison (steps, seconds/image, GPU-days) remains valid independently.
minor comments (8)
  1. [Section 2.1, Eq. (4)] The statement that the temperature τ is 'set to 0' during inference is not literally meaningful in Eq. (4) because τ divides the logits; please clarify that the zero-temperature limit or a deterministic threshold is intended, and specify the straight-through estimator used for Bernoulli sampling during training.
  2. [Section 2.1 and Section 4.1] There are multiple typos that should be fixed: 'calculat' should be 'calculate', 'deocder' should be 'decoder', 'Notebaly' should be 'Notably', and the Introduction contains 'avarage' and 'scaore'.
  3. [Appendix C, Algorithm 2] The algorithm title reads 'Samplified sampling algorithm' and should be 'Simplified sampling algorithm'.
  4. [Appendix D.2] The FID protocol says 'MSCOCO val2024 set' but the comparison set is 'MSCOCO val2014'; please correct the year and make the number of generated images (3K vs 30K) consistent with the Table 3 column label.
  5. [Table 3 and Appendix D.1] Appendix D.1 states that all Table 3 results use 20 sampling steps with Algorithm 2, but Table 3 also reports a 50-step row; please specify the configuration used for that row.
  6. [Table 2] BIT is trained on LAION-COCO rather than ImageNet, so it should carry the † marker used in the table for other out-of-domain tokenizers; without the marker, the comparison to ImageNet-trained TiTok-B and LlamaGen appears to be a same-domain comparison.
  7. [Table 1 and Section 4] Model sizes are inconsistent: the text says the diffusion and AR models are 1B total parameters, while Table 1 lists Instella AR as 0.8B and Instella Diff as 1.2B, and Table 3 says 'Diffusion BIT 1B'; please reconcile these numbers.
  8. [Quantitative tables] All quantitative tables report point estimates without error bars or multiple evaluation seeds; for GenEval and ImageReward, differences of a few hundredths are common, and stating the variance or seed count would make the comparisons more reliable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims rest on external benchmarks and the cited binary-latent diffusion component is an independent prior publication, not a premise that forces the present results.

full rationale

The paper's headline claims — competitive text-to-image performance with 128 discrete tokens, training within 200 GPU-days, and sub-second inference — are evaluated against external baselines and external benchmarks. Table 1 compares GenEval, CLIP, and ImageReward scores against SD, PixArt, SD3, Chameleon, and Emu3, whose published numbers were not produced by this paper. Table 2 compares the tokenizer's rFID, PSNR, and SSIM against SD-VAE, DC-AE, TiTok, and VQ-GAN. Table 3 compares FID, seconds per image, and GPU-days against SDv1.5 and PixArt. These are independent, externally falsifiable measurements. The binary latent diffusion component is cited from Wang et al. (2023a), which shares an author (Zicheng Liu) with the present work, so this is a self-citation; however, that prior work is an externally published CVPR paper providing a modeling primitive (Bernoulli diffusion), and the present paper extends it from discrete to continuous time and applies it to a new 1D tokenizer. None of the present paper's predictions are defined in terms of the fitted tokenizer or the cited prior work in a way that forces the reported numbers. The token count of 128 is an architectural choice, not a fitted parameter. Appendix D.2's caveat about 'potential mismatches about the detailed evaluation protocols across previous works' is an honest comparability limitation, not a circular reduction: it affects whether the baseline numbers are directly comparable, but it does not make the present model's scores equivalent to its inputs by construction. No equation in the paper defines a predicted metric as the same quantity used in a fit, and no load-bearing uniqueness or ansatz is smuggled in only via self-citation. The derivation chain is therefore self-contained with respect to the claims evaluated.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central claim rests on hyperparameters (token count, binary dimension, inference settings) chosen by hand, on the Bernoulli diffusion process imported from prior work, and on assumptions about resolution generalization, text conditioning, data sufficiency, and evaluation comparability. No new physical entity is introduced; the 1D binary latent is a designed representation whose performance is measured only inside this paper.

free parameters (6)
  • Number of latent tokens k = 128 (with 64 binary bits each)
    Set by hand as the headline 128-token representation; the token count reduction claim is exactly this number.
  • Binary vector dimension c_hat = 64
    Chosen together with k so each image is 8192 binary bits; balances capacity against sequence length.
  • Patch size p = 16x16
    Chosen for efficiency; Appendix E notes patch size 8 improves reconstruction but is too costly.
  • Sampling temperature for Bernoulli quantization during training = tau=0.5 (stage 1), tau=0.1 (stage 2)
    Hand-tuned to prevent code collapse while retaining robustness; affects reconstruction and downstream generation.
  • CFG scale alpha and temperature tau at inference = alpha=7.5, tau=0.75 (Table 1); alpha=4.0, tau=0.6 (Table 3); alpha=7.5-9.0, tau=0.6-1.0 (figures)
    These are chosen per benchmark and influence every reported GenEval, CLIP, ImageReward, and FID number, so the scores are not parameter-free.
  • Number of inference denoising steps = 100 (Table 1), 20 and 50 (Table 3)
    The sub-second claim uses 20 steps; FID improves from 16.33 to 15.10 when going from 20 to 50 steps, showing sensitivity to this choice.
assumptions (6)
  • domain assumption The Bernoulli diffusion forward process z_t ~ B(0.5t + (1-t)z) is a valid generative model for binary latents.
    Adopted from Wang et al. (2023a), Section 2.2. The T2I diffusion training and sampling depend on this process, including the BCE loss and CFG formula.
  • domain assumption Bernoulli sampling in the quantizer with temperature annealing prevents code collapse and yields useful discrete latents.
    Section 2.1 "Binary quantizer" asserts this empirically; reconstruction and generation quality rest on it.
  • domain assumption The decoder generalizes to 768x768 and 1024x1024 even though the encoder is only trained with 512x512 inputs in stage 2.
    Section 2.1 "Multi-stage training" specifies encoder at 512 while decoder sees multiple resolutions; the 1024x1024 generation claim relies on this generalization.
  • domain assumption Frozen AMD OLMo-1B intermediate-layer text features are sufficient text conditioning for T2I.
    Section 4 "Model Architectures"; the whole text-image alignment depends on this choice, with no comparison against CLIP/T5 encoders.
  • domain assumption The evaluation protocols in Table 1 are comparable to previously reported numbers for other models.
    Appendix D.2 admits potential mismatches across works; the "competitive" claim depends on this comparability.
  • domain assumption Public LAION and synthetic data suffice to train a competitive T2I model.
    Appendix B.1 lists data sources; the absence of private data claim depends on this.
invented entities (1)
  • 1D binary image latent
    purpose: Compact image representation: 128 binary vectors of dimension 64 replace one-hot codebook tokens for T2I generation.
    The representation is a new design introduced here; its performance is only reported in this paper's own tables, so there is no external handle. Its reconstruction and generation metrics are internally generated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Instella-T2I: Pushing the Limits of 1D Discrete Latent Space Image Generation." pith.science (2026). https://pith.science/paper/3NLVTSSA

@misc{pith2026250621022,
  author       = {Pith},
  title        = {Pith review of: Instella-T2I: Pushing the Limits of 1D Discrete Latent Space Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3NLVTSSA}},
  note         = {Machine review of arXiv:2506.21022}
}
read the original abstract

Image tokenization plays a critical role in reducing the computational demands of modeling high-resolution images, significantly improving the efficiency of image and multimodal understanding and generation. Recent advances in 1D latent spaces have reduced the number of tokens required by eliminating the need for a 2D grid structure. In this paper, we further advance compact discrete image representation by introducing 1D binary image latents. By representing each image as a sequence of binary vectors, rather than using traditional one-hot codebook tokens, our approach preserves high-resolution details while maintaining the compactness of 1D latents. To the best of our knowledge, our text-to-image models are the first to achieve competitive performance in both diffusion and auto-regressive generation using just 128 discrete tokens for images up to 1024x1024, demonstrating up to a 32-fold reduction in token numbers compared to standard VQ-VAEs. The proposed 1D binary latent space, coupled with simple model architectures, achieves marked improvements in speed training and inference speed. Our text-to-image models allow for a global batch size of 4096 on a single GPU node with 8 AMD MI300X GPUs, and the training can be completed within 200 GPU days. Our models achieve competitive performance compared to modern image generation models without any in-house private training data or post-training refinements, offering a scalable and efficient alternative to conventional tokenization methods.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond Patches: Global-aware Autoregressive Model for Multimodal Few-Shot Font Generation

    cs.CV 2026-01 unverdicted novelty 7.0 of 10

    GAR-Font is a global-aware autoregressive framework for multimodal few-shot font generation that adds global tokenization, a language-style adapter, and post-refinement to improve style coherence over patch-based methods.

  2. VibeToken: Scaling 1D Image Tokenizers and Autoregressive Models for Dynamic Resolution Generations

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    VibeToken enables autoregressive image generation at arbitrary resolutions using 64 tokens for 1024x1024 images with 3.94 gFID, constant 179G FLOPs, and better efficiency than diffusion or fixed AR baselines.

Reference graph

Works this paper leans on

30 extracted references · 3 canonical work pages · cited by 2 Pith papers

  1. [1]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901,

    1901

  2. [6]

    Fifty, R

    C. Fifty, R. G. Junkins, D. Duan, A. Iger, J. W. Liu, E. Amid, S. Thrun, and C. Ré. Restructuring vector quantization with the rotation trick.arXiv preprint arXiv:2410.06424,

    arXiv

  3. [7]

    URL https://arxiv.org/abs/2307.08041. D. Ghosh, H. Hajishirzi, and L. Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment.Advances in Neural Information Processing Systems, 36,

    arXiv

  4. [9]

    Hessel, A

    J. Hessel, A. Holtzman, M. Forbes, R. L. Bras, and Y. Choi. Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718,

    arXiv

  5. [12]

    D. Kim, J. He, Q. Yu, C. Yang, X. Shen, S. Kwak, and L.-C. Chen. Democratizing text-to-image masked generative models with compact text-aware one-dimensional tokens. arXiv preprint arXiv:2501.07730,

    arXiv

  6. [14]

    URLhttps://arxiv.org/abs/ 2211.09117. T. Li, Y. Tian, H. Li, M. Deng, and K. He. Autoregressive image generation without vector quantization. NeurIPS, 2024a. X. Li, H. Chen, K. Qiu, J. Kuen, J. Gu, B. Raj, and Z. Lin. Imagefolder: Autoregressive image generation with folded tokens.arXiv preprint arXiv:2410.01756, 2024b. N. Ma, M. Goldstein, M. S. Albergo,...

    work page Pith review arXiv

  7. [15]

    org/abs/2212.09748

    URLhttps://arxiv. org/abs/2212.09748. 13 Instella-T2I: Pushing the Limits of 1D Discrete Latent Space Image Generation D. Podell, Z. English, K. Lacey, A. Blattmann, and T. Dockhorn. Sdxl: Improving latent diffusion models for high-resolution image synthesis.ICLR,

    arXiv

  8. [16]

    L. Qu, H. Zhang, Y. Liu, X. Wang, Y. Jiang, Y. Gao, H. Ye, D. K. Du, Z. Yuan, and X. Wu. To- kenflow: Unified image tokenizer for multimodal understanding and generation.arXiv preprint arXiv:2412.03069,

    arXiv

Show all 30 references
  1. [18]

    Takida, T

    Y. Takida, T. Shibuya, W. Liao, C.-H. Lai, J. Ohmura, T. Uesaka, N. Murata, S. Takahashi, T. Kumakura, and Y. Mitsufuji. Sq-vae: Variational bayes on discrete representation with self-annealed stochastic quantization. arXiv preprint arXiv:2205.07547,

    arXiv

  2. [19]

    C. Team. Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint arXiv:2405.09818,

    arXiv

  3. [20]

    URLhttps://arxiv.org/abs/2404.02905. H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhar- gava, S. Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

    arXiv

  4. [22]

    Z. Wang, J. Wang, Z. Liu, and Q. Qiu. Binary latent diffusion. InCVPR, pages 22576–22585, 2023a. Z. Wang, J. Wang, Z. Liu, and Q. Qiu. Binary latent diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22576–22585, 2023b. Z. J. ...

    arXiv

  5. [23]

    J. Wu, Y. Jiang, C. Ma, Y. Liu, H. Zhao, Z. Yuan, S. Bai, and X. Bai. Liquid: Language models are scalable multi-modal generators.arXiv preprint arXiv:2412.04332,

    arXiv

  6. [24]

    J. Yu, X. Li, J. Y. Koh, H. Zhang, R. Pang, J. Qin, A. Ku, Y. Xu, J. Baldridge, and Y. Wu. Vector-quantized image modeling with improved vqgan.arXiv preprint arXiv:2110.04627,

    arXiv

  7. [25]

    L. Yu, J. Lezama, N. B. Gundavarapu, L. Versari, K. Sohn, D. Minnen, Y. Cheng, A. Gupta, X. Gu, A. G. Hauptmann, B. Gong, M.-H. Yang, I. Essa, D. A. Ross, and L. Jiang. Language model beats diffusion - tokenizer is key to visual generation. InThe Twelfth International Conferen...

    arXiv

  8. [26]

    URLhttps://arxiv.org/abs/1801.03924. Y. Zhao, Y. Xiong, and P. Krähenbühl. Image and video tokenization with binary spherical quantization. arXiv preprint arXiv:2406.07548,

    arXiv

  9. [27]

    Y. Zhu, B. Li, Y. Xin, and L. Xu. Addressing representation collapse in vector quantized models with one linear layer.arXiv preprint arXiv:2411.02038,

  10. [28]

    Zhuang, Y

    X. Zhuang, Y. Xie, Y. Deng, L. Liang, J. Ru, Y. Yin, and Y. Zou. Vargpt: Unified understanding and gener- ation in a visual autoregressive multimodal large language model.arXiv preprint arXiv:2501.12327,

    arXiv

  11. [29]

    15 Instella-T2I: Pushing the Limits of 1D Discrete Latent Space Image Generation A. Text-to-Image Prompts The prompts, model, and inference configurations used for the text-to-image samples demonstrated in Figure 1 are listed below: • A fashion style pretty woman portrait wear...

    2024

  12. [256]

    FID score

    The final reported number is an average over all the 30K samples. FID score. For text-to-image model evaluation, we adopt the evaluation code of Kang et al. (2023). The images are generated using 3K text prompts from the MSCOCO val2024 set, with one image per prompt. The FID i...

    2023

  13. [2016]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Min- derer, G.Heigold, S.Gelly, etal. Animageisworth16x16words: Transformersforimagerecognition at scale.arXiv preprint arXiv:2010.11929,

    2010 arXiv

  14. [2017]

    X. Wang, X. Zhang, Z. Luo, Q. Sun, Y. Cui, J. Wang, F. Zhang, Y. Wang, Z. Li, Q. Yu, et al. Emu3: Next-token prediction is all you need.arXiv preprint arXiv:2409.18869,

    arXiv

  15. [2018]

    URLhttps://arxiv.org/abs/1611.07004. J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pag...

    2016 arXiv

  16. [2019]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022a. R. Rombach, A. Blattmann, D. Lorenz, P. Ess...

    arXiv

  17. [2020]

    J. Han, J. Liu, Y. Jiang, B. Yan, Y. Zhang, Z. Yuan, B. Peng, and X. Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis.arXiv preprint arXiv:2412.04431,

    arXiv

  18. [2021]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    arXiv

  19. [2022]

    URLhttps://arxiv.org/abs/2202.04200. H. Chang, H. Zhang, J. Barber, A. Maschinot, J. Lezama, L. Jiang, M.-H. Yang, K. Murphy, W. T. Freeman, M. Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704,

    arXiv

  20. [2023]

    11 Instella-T2I: Pushing the Limits of 1D Discrete Latent Space Image Generation H. Chen, Z. Wang, X. Li, X. Sun, F. Chen, J. Liu, J. Wang, B. Raj, Z. Liu, and E. Barsoum. Softvq-vae: Efficient 1-dimensional continuous tokenizer.arXiv preprint arXiv:2412.10958, 2024a. H. Chen,...

    arXiv

  21. [2024]

    L. Fan, T. Li, S. Qin, Y. Li, C. Sun, M. Rubinstein, D. Sun, K. He, and Y. Tian. Fluid: Scaling autore- gressive text-to-image generative models with continuous tokens.arXiv preprint arXiv:2410.13863,

    arXiv

  22. [2025]

    D. P. Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.