Introduction
The paper models continuous-token distributions with diffusion, connecting autoregression to continuous latents and showing that autoregressive image generation need not require discrete quantisation.
View full image ↗Figure 1: Diffusion Loss for continuous tokens and autoregressive conditioning. Source: v3; the image version is distinct from the initial submission date.
Li, Tianhong; Tian, Yonglong; Li, He; Deng, Mingyang; He, Kaiming · original paper / arXiv · Rights in the paper and depicted works remain with their holders. Research quotation does not establish an open licence; republication rights await independent review.
Source · Autoregressive Image Generation without Vector Quantization ↗RESEARCH ACCOUNT
The paper models continuous-token distributions with diffusion, connecting autoregression to continuous latents and showing that autoregressive image generation need not require discrete quantisation.
ANTI-AI ARCHIVE · Revised 2026-10-03
What the paper investigates
Diffusion Loss replaces the usual categorical cross-entropy in modelling per-token distributions. Experiments cover standard autoregression and masked autoregressive variants. Diffusion handles local probability modelling while sequence organisation remains a separate design choice.
Section sources: Autoregressive Image Generation without Vector Quantization
Its place in generative-art history
Editorial interpretation: a binary choice between diffusion and autoregression cannot describe every tool. This hybrid approach makes it possible to track encoding, generation order and sampling as distinct design decisions.
Section sources: Autoregressive Image Generation without Vector Quantization
Reading limits and versions
Continuous representations remain constrained by the encoder and training data; removing quantisation does not imply lossless preservation of image information.
Section sources: Autoregressive Image Generation without Vector Quantization
Sources for this account
Read the arXiv abstract, authors and version history, plus the selected figure/page on PDF page 2; proofs and full experiments were not independently audited.
Read the arXiv abstract, authors and version history, plus the selected figure/page on PDF page 2; proofs and full experiments were not independently audited. The account and translation are AI-assisted, pending independent human review. Section references identify evidence without claiming independent verification of every historical statement.
Continue with a comparative question
MAR / Diffusion Loss: compare methods, control or evaluation conditions with VAR.
MAR / Diffusion Loss: compare methods, control or evaluation conditions with Scaling Autoregressive Models for Content-Rich Text-to-Image Generation.
MAR / Diffusion Loss: compare methods, control or evaluation conditions with PixArt-α.
Dates & version record
Date displayed for this node: 2024-06-17 · Historical date recorded for the source: 2024-06-17
The timeline uses the initial arXiv submission, distinct from conference publication, model release and the archive addition on 3 October 2026. The account and image refer to v3; its date is recorded in the source history.
These dates refer to the historical event or recorded version, not this page’s publication date. The original date precision and unresolved questions are retained.
Original sources & further reading
These links lead to the cited paper, article, institution or conference page. External texts retain their source languages.
Autoregressive Image Generation without Vector Quantization
https://arxiv.org/abs/2406.11838
Current cited URL
Read the arXiv abstract, authors and version history, plus the selected figure/page on PDF page 2; proofs and full experiments were not independently audited.
Provenance, translation & verification
Archive node #279 · Initial research-paper submission; methods, evaluation and production conditions
Research materials & supplement references · 1
Recent research papers: additions and reconciled sources since 2022 · Read the arXiv abstract, authors and version history, plus the selected figure/page on PDF page 2; proofs and full experiments were not independently audited. · 026
Date as recorded: 2024-06-17
Autoregressive Image Generation without Vector Quantization ↗
Autoregressive Image Generation without Vector Quantization
Archive account based on the listed research materials, not a full translation of the linked work. AI-assisted translation; independent human review pending.
Official material was read within the stated scope; see the source note for reading limits and outstanding checks.
Cite this node
ANTI-AI ARCHIVE. Autoregressive Image Generation without Vector Quantization. Art-history node #279. https://salondesrefuses.cn/en/art-history/349
For specific historical claims, also cite the original sources above and include your access date. This account is not a full translation of the linked work.
Adjacent nodes follow chronological order; adjacency does not establish direct influence or causation.