System One models such as Jev return decisions through fast, parallel prediction. In this paper, we introduce Jeveration, showing that the same decision paradigm can also generate high-quality images in a single forward pass. Jeveration builds on a frozen vision-language model, repurposing its visual-semantic knowledge for holistic image generation. The key challenge is to maintain global consistency. A single prompt admits many valid images, yet parallel predictions must converge on a coherent whole. We therefore condition all positions on a shared random state to coordinate parallel predictions, so that each draw yields a different but coherent image. Since the model tends to ignore this state under direct image supervision, we first learn a distribution over discrete image codes to guide single-pass generation. Beyond text-to-image generation, we further unify image-to-image generation and dense prediction within the same architecture, so that one checkpoint supports all three tasks, each in one forward pass. Experiments show that Jeveration generates a 512x512 image in 78 ms, up to 122x faster than existing unified models, while maintaining competitive image quality and dense prediction performance.
Figures & tables
Figure 2: Direct parallel selection yields incoherent images. Codes sampled independently from the first-phase model do not fit together, while Jeveration produces coherent images.
Figure 3: Architecture overview. A frozen vision-language model with shared LoRA adapters processes input and decision tokens in one forward pass. Decision tokens start from a random state for image generation, while for dense prediction they act as spatial queries. Task-specific heads then map their final hidden states to image codes or dense predictions. One checkpoint serves all tasks.
Figure 4: Text-to-image generation. Samples of the final Jeveration checkpoint at 512 px, with excerpts of their prompts.
Model
Unified
Params (B)
Steps
Res. (px)
Speedup ↑
GenEval ↑
DPG ↑
HPSv2.1 ↑
Latency (ms) ↓
Generation-only models
PixArt- α [ 8 ]
×
0.6
20
512
25.1 ×
0.49
71.6 ‡
29.98
378.8
PixArt- Σ [ 7 ]
×
0.6
20
512
22.7 ×
0.54
79.5 ‡
30.49
419.5
SANA-0.6B [ 48 ]
×
0.6
20
512
16.2 ×
0.65
84.3 †
29.86
587.3
SANA-Sprint 0.6B [ 10 ]
×
0.6
1
1024
78.5 ×
0.72
78.63
28.74
121.2
SDXL-DMD2 [ 52 ]
×
2.6
1
1024
46.6 ×
0.56
74.09
28.20
204.0
Table 1: Text-to-image quality and end-to-end latency. All models are timed on the same H200 at batch 1, including decoding, with released samplers at the listed steps and resolutions. Speedup is relative to Janus-Pro-7B, the slowest model listed. Params include our frozen backbone; Steps count sequential generator evaluations, or tokens for autoregressive models. Colored squares identify image-generation families, including distilled variants. † / ‡ indicate scores from the authors / another paper.
Figure 5: Image-to-image generation on held-out examples. In each pair the left image is the input and the right image the output.
Structure follow ↑
Style
Model
HED
Depth
Colorize
PSNR ↑
Jeveration
0.880
0.802
0.937
15.40
Copy the input
0.828
0.713
0.998
–
Decoder without source
0.792
0.767
0.757
14.37
Table 2: Image-to-image generation. Structure fidelity on 341 held-out examples per task and style PSNR against paired targets.
Seg.
Depth
Normal
Model
mIoU ↑
abs. (m) ↓
δ1↑
mean ( ∘ ) ↓
Jeveration
67.7
0.267
0.898
24.2
Dense-only, one task each
67.8
0.246
0.915
21.2
Dense-only, three tasks
67.3
0.253
0.912
23.0
DeepLabV3+ R50 ‡ [ 31 , 30 ]
53.5
0.393
–
22.0
Table 3: Dense prediction on NYUv2. Dense-only models use the same backbone. ‡ is a reported result.
Figure 6: Dense prediction of 13-class segmentation, depth and surface normals on NYUv2 test pictures with the same checkpoint.
Variant
GenEval ↑
HPS ↑
w/ random state
0.73
28.27
w/ fixed state
0.70
26.63
w/ correction target
0.74
28.03
w/o correction target
0.73
27.92
Image head (wide + codebook)
0.64
25.42
Image head (wide only)
0.60
25.34
Table 4: Ablation studies. Each block is a separate experiment that changes one component, so values are comparable only within a block. From top to bottom, the blocks start from a text-to-image-only model, an earlier run on all tasks and the first-phase masked model. Highlighted rows are the settings Jeveration uses.
Training tasks
GenEval ↑
HPS ↑
DPG ↑
Text-to-image only
0.73
28.28
78.61
All tasks (Jeveration)
0.74
28.06
78.24
Table 5: Joint training across tasks. Training on all tasks costs almost no text-to-image quality. Both models get about the same amount of text-to-image training, so the difference comes from adding image-to-image and dense tasks.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
Text-to-image generation has advanced rapidly with diffusion models, progressing from CLIP and T5 conditioning to unified systems where a single LLM backbone handles both visual understanding and generation. Despite the architectural unification, these systems frequently fail to faithfully align complex prompts during synthesis, even though they remain highly accurate at verifying whether an image satisfies those same prompts. We formalize this as the understanding-generation gap and propose UniReasoner, a framework that leverages the LLM as a universal reasoner to convert its understanding strength into direct generation guidance. Given a prompt, the LLM first produces a coarse visual draft composed of discrete vision tokens. It then performs a self-critique by evaluating the draft for prompt consistency, producing a grounded textual evaluation that pinpoints what needs to be corrected. Finally, a diffusion model is conditioned jointly on the prompt, the visual draft, and the evaluation, ensuring that generation is guided by explicit corrective signals. Each signal addresses a limitation of the other: the draft provides a concrete, scene-level anchor that reduces under-specification in text-only conditioning, while the evaluation turns verification into grounded, actionable constraints that correct omissions, hallucinations, and relational errors. Experiments show that UniReasoner improves compositional alignment and semantic faithfulness under the same diffusion backbone while maintaining image quality, demonstrating a practical way to exploit LLM reasoning to close the understanding-generation gap.
We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior through image-only pre-training and mid-training. The generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image into LLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.
Instruction-based image editing models exhibit pronounced sample-specific performance variation: different models can produce substantially different results for the same source image and editing instruction, and the best-performing model often varies across requests. These models also differ substantially in inference latency, making fixed-model deployment inefficient. We therefore study image editing model routing, which dynamically selects an editor from a candidate pool given an input image and editing instruction. We introduce EditJev, a System One routing model that directly predicts sample-level preference scores over candidate image editing models in a single forward pass. A lightweight preference head on a vision-language backbone produces a fixed-dimensional score vector without autoregressive model-name generation or response parsing. To train EditJev, we construct pairwise preferences from offline candidate outputs and optimize a pairwise ranking objective to learn relative model suitability. At inference time, EditJev combines the predicted preference scores with model-level speed scores to balance editing quality and generation latency. Experiments on ImgEdit and GEdit-En show that EditJev improves aggregate quality over the strongest fixed editor, from 4.540 to 4.592 and from 8.194 to 8.255, respectively, while reducing estimated end-to-end latency by 60.8% and 31.7%. These results demonstrate the effectiveness of direct, structured model routing for efficient image editing.
Zhile Guan, Xu Jiang, Yule Duan +4
School of Electronic and Computer Engineering, Peking University · Tsinghua University