OneFlow: Concurrent Mixed-Modal and Interleaved Generation with Edit Flows

TL;DR

OneFlow is the first non-autoregressive multimodal model that generates text and images concurrently in a single transformer, combining an insertion-based Edit Flow for discrete text tokens with Flow Matching for continuous image latents. Unlike Transfusion (AR text + diffusion image), there is no causal ordering between modalities — text and images can be inserted anywhere in the sequence and refined in parallel under a hierarchical sampling schedule that prioritizes content over grammar. Scales better than Transfusion at matched compute (1B–8B controlled comparisons) and matches or beats it with up to 50% fewer training FLOPs, while unlocking variable-length output and per-image time schedules. The two emergent capabilities worth flagging: classifier-free guidance can be applied to image understanding (longer, more detailed captions at higher CFG scales), and the hierarchical sampler exhibits implicit reasoning-like behavior on VQA without CoT prompting.

Key claims

OneFlow combines Edit Flow (insertion-based discrete generation for text) with Flow Matching (continuous generation for image latents) under a shared transformer backbone with U-Net projectors, so text tokens and image embeddings live in one sequence and are refined concurrently [§1, §3].
During training, Edit Flow deletes tokens/images to shorten sequences (vs masked diffusion, which replaces with mask tokens at the same length), saving FLOPs at the same target sequence — the primary mechanism behind the 50% training-FLOP reduction vs Transfusion [§3, Fig. on training mechanics].
In controlled experiments from 1B to 8B, OneFlow exhibits strictly better scaling laws than the AR+FM Transfusion baseline on both image generation (DPG-Bench, FID) and image-to-text understanding (CIDEr, ROUGE) [Abstract, §scaling experiments].
Mixed-modal pretraining (text and images jointly) outperforms sequential single-modality pretraining: +4% relative on VQA tasks plus a slight image-generation lift, evidence that the two objectives mutually reinforce when concurrent generation is possible [§ablation on mixed vs sequential pretraining].
Classifier-free guidance can be applied on the image-understanding side: higher CFG values during caption generation consistently produce longer, more detailed text, at the cost of higher hallucination risk — a capability that has no natural analog in AR-based UMMs because there is no separate unconditional branch over the text [§CFG analysis].
Hierarchical sampling (per-modality, per-image time schedule) prioritizes content over grammar and produces implicit reasoning chains during VQA without explicit chain-of-thought prompting — interpreted as the sampler naturally drafting structured intermediate states before final commitment [§Hierarchical Generation].
OneFlow generates variable-length outputs natively via insertion + deletion; this contrasts with masked diffusion models which are fixed-length, and with AR which is one-token-at-a-time sequential. Variable length plus concurrent generation enables interleaved text + variable-number-of-images outputs that prior unified models cannot produce in one pass [§1, Fig. 1].

Method

OneFlow factors a multimodal sequence into an Edit-Flow process over discrete text tokens and a Flow-Matching process over continuous image latents, both parameterized by one shared transformer with U-Net projectors at the image boundaries (the same architectural slot Transfusion uses). The training objective is the sum of the Edit-Flow loss (insertion / deletion targets on text) and the Flow-Matching velocity loss (on image latents), with each image carrying its own time variable independent of the text time and other images’ times.

At training, the model sees mixed sequences where text tokens and image embeddings have been deleted (Edit Flow) and noised (Flow Matching) according to per-element schedules; the model learns to insert text tokens at the right positions and denoise image latents in parallel. At inference, the sampler maintains independent time variables per modality element and uses a hierarchical schedule that resolves coarse content (insertion structure, low-noise image content) before fine detail (grammatical tokens, high-noise image content). Because deletion shortens the sequence, training-time FLOPs per token are lower than masked diffusion’s at the same end length; this is the source of the 50%-FLOP-savings claim against Transfusion at matched quality.

Results

Across 1B–8B model sizes, OneFlow outperforms a like-for-like Transfusion (AR text + FM image) baseline on DPG-Bench, FID, CIDEr, and ROUGE under matched training tokens, with the gap widening with scale (better scaling exponent) [§scaling experiments].
At matched performance, OneFlow uses up to 50% fewer training FLOPs than Transfusion [Abstract].
Mixed-modal pretraining vs sequential pretraining: +4% relative VQA improvement, with slight image-generation gains as well [§ablation].
512×512 T2I samples qualitatively follow fine-grained prompt details (e.g. “polar bear balancing on a blue barrel” gets the relational structure right) [Figure: generated samples].
CFG on the understanding side: monotonically longer / more detailed captions as CFG weight increases, with documented hallucination tradeoff [§CFG analysis, qualitative].
Competitive with existing unified-model SOTA on standard image-generation and image-understanding benchmarks at 8B [Abstract, Table comparing to existing UMMs].

Why it’s interesting

The wiki has been tracking the unified-multimodal-models space as a three-way split — AR (Show-o/Janus), AR+Diffusion (Transfusion, BAGEL, DuoGen: Towards General Purpose Interleaved Multimodal Generation), and tri-modal masked diffusion (The Design Space of Tri-Modal Masked Diffusion Models). OneFlow opens a fourth corner: Flow+Flow — non-AR on both sides, where the discrete side is Edit Flow rather than masked diffusion. This matters because Edit Flow’s deletion primitive gives variable-length output natively, which the fixed-length masked-diffusion side has been the perennial weakness of (e.g. the tri-modal MDM design space’s question about how to handle variable-length generation). OneFlow is also the cleanest demonstration on the wiki that “mixed-modal pretraining beats sequential pretraining” — a hypothesis the UMM cluster has been collecting partial evidence for (UniMRG’s training-time auxiliary generation in Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation is a different recipe for the same hypothesis).

Two cross-cutting connections worth flagging: (a) the CFG-on-understanding observation is a genuinely new axis for Classifier-Free Guidance variants — every prior filed CFG variant operates on the generation side, and OneFlow shows the conditional/unconditional branch slot exists naturally for caption generation when the model is non-AR. (b) The hierarchical-sampler-as-implicit-reasoner observation parallels DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models‘s “diffusion sampling is iterative reasoning” framing — both suggest that the iterative refinement structure of flow/diffusion samplers is doing something AR models pay extra for via CoT prompting.