Jeveration

System One Models Now Generate

A frozen vision-language model returns a whole 512 px image as one decision, in a single forward pass.

Yuhe Liu*, Qi Li*, Shihua Zhang, Xinchao Wang†

National University of Singapore
*Equal contribution · †Corresponding author

Text-to-image samples around a mosaic of many generated thumbnails that spells Jeveration, with image-to-image results and NYUv2 dense predictions below.
Text-to-image samples (top), image-to-image generation and dense prediction (bottom). One checkpoint, one forward pass per task.

Highlights

78 ms
for a 512 px image, end to end on one H200
122×
faster than Janus-Pro-7B, at 0.74 GenEval
1 checkpoint
for text-to-image, image-to-image and dense prediction

Latency per image

End to end on one H200 at batch 1, decoding included

Abstract

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 512×512 image in 78 ms, up to 122× faster than existing unified models, while maintaining competitive image quality and dense prediction performance.

How it works

Generation is one decision

Autoregressive models generate an image token by token, and diffusion models by iterative denoising. Jeveration removes the sequential sampling. It asks a frozen vision-language model for one typed decision, in which all 128 codes of the image are chosen in parallel.

Parallel choices need a shared random state

One prompt admits many valid images. When every code is chosen on its own, different positions follow different images and the regions do not fit together. Jeveration draws one random state per image and conditions every position on it, so each draw yields a different but coherent image.

Portrait with no joint decision, its regions do not match.
No joint decision
Portrait of a woman wearing a flower crown generated by Jeveration.
Jeveration
Snow leopard with no joint decision, its regions do not match.
No joint decision
Snow leopard resting on a rock ledge generated by Jeveration.
Jeveration

Codes sampled independently from the first-phase model do not fit together, while Jeveration produces coherent images.

Learned in two phases

Trained on target images alone, the model ignores the random state. Training therefore has two phases.

  1. Learn the image distribution

    The model first learns to fill in masked image codes. This teaches it how the codes of an image depend on each other.

  2. Learn to decide

    The model then turns random states into whole images in one pass. The first-phase model judges each sample and guides the decisions until the codes fit together.

The first-phase model is used only in training, so inference stays a single forward pass.

One frozen model for every task

The backbone is a frozen Qwen3.5-4B with shared LoRA adapters. Task heads read the decision tokens and return image codes, or segmentation, depth and surface normals for dense prediction.

Architecture: input and decision tokens pass through a frozen vision-language model with shared LoRA adapters in one forward pass. An image head and decoder produce text-to-image and image-to-image outputs, and dense heads produce segmentation, depth and normals.
Input and decision tokens pass through the frozen model once. Decision tokens start from a random state for image generation and act as spatial queries for dense prediction.

Results

One checkpoint, three tasks

Text-to-image

Selected 512 px samples, each from one forward pass. Select a picture to see its prompt.

Quality and latency

Jeveration is the fastest model in the comparison at 77.8 ms and reaches 0.74 GenEval. It is 98× and 122× faster than Janus-Pro-1B and Janus-Pro-7B, and scores 0.11 higher GenEval than four-step Show-o Turbo-512 at the same resolution.

GenEval against latency per image

Batch 1 on one H200, decoding included. Latency on a log scale.

  • Jeveration (ours)
  • Unified models
  • Generation-only models
Full comparison table
ModelFamilyUnifiedParams (B)StepsRes. (px)Speedup ↑GenEval ↑DPG ↑HPSv2.1 ↑Latency (ms) ↓

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. Family names the image-generation family, including distilled variants. † and ‡ mark scores from the authors and from another paper.

Image-to-image

The same checkpoint turns edge maps and depth maps into images, colorizes grayscale photos and transfers styles. A request takes 91.7 ms, only 14 ms more than text-to-image. Point at a picture or drag across it to compare input and output.

Output: a green jeep parked on a rocky mountain cliff.Input: soft HED edge map of a jeep on a cliff.InputOutput
HED edges → image
Output: an oval diamond ring.Input: depth map of a ring, bright is near.InputOutput
Depth map → image
Output: the mug in color, with a red flower and green leaves.Input: grayscale photo of a mug with a flower print.InputOutput
Colorization
Output: the pizza slice as an anime illustration.Input: photo of a pizza slice on a plate.InputOutput
Anime
Output: the earrings in Studio Ghibli style with a sky background.Input: photo of a pair of hoop earrings.InputOutput
Studio Ghibli
Output: the small airplane as a watercolor painting.Input: photo of a small airplane on a runway.InputOutput
Watercolor
Scores
ModelHED ↑Depth ↑Colorize ↑Style PSNR ↑
Jeveration0.8800.8020.93715.40
Copy the input0.8280.7130.998–
Decoder without source0.7920.7670.75714.37

Structure tasks also receive a caption, while style tasks use a fixed instruction for each style. Structure fidelity on 341 held-out examples per task (HED and colorization by edge F1, depth by δ1) and style PSNR against paired targets.

Dense prediction

The same checkpoint predicts segmentation, depth and surface normals on NYUv2 in about 78 ms. It reaches 67.7 mIoU, on par with dense-only models trained from the same backbone.

NYUv2 test picture of a living room with a red sofa.
Input
Predicted 13-class segmentation.
Segmentation
Predicted depth, near is bright.
Depth
Predicted surface normals.
Normals
NYUv2 test picture of a bedroom.
Input
Predicted 13-class segmentation.
Segmentation
Predicted depth, near is bright.
Depth
Predicted surface normals.
Normals
NYUv2 test picture of a bedroom with framed pictures and a lamp.
Input
Predicted 13-class segmentation.
Segmentation
Predicted depth, near is bright.
Depth
Predicted surface normals.
Normals
Scores
ModelSeg. mIoU ↑Depth abs. (m) ↓Depth δ1 ↑Normal mean (°) ↓
Jeveration67.70.2670.89824.2
Dense-only, one task each67.80.2460.91521.2
Dense-only, three tasks67.30.2530.91223.0
DeepLabV3+ R50‡53.50.393–22.0

NYUv2, 654 test images. Dense-only models use the same backbone. ‡ is a reported result.

Ablations

The effect of each component

The correction target brings a further gain in image quality, and the two branches of the image head together outperform either branch alone. Training on all tasks costs almost no text-to-image quality.

Correction target

Removing the correction target lowers HPSv2.1 at similar GenEval. This suggests that distribution matching sets the composition, while the correction refines details.

VariantGenEvalHPSv2.1
With correction0.7428.03
Without correction0.7327.92

Image-head branches

The image head predicts each code with two branches, one initialised from the tokenizer's codebook and a wider one that learns freely. Together they outperform either branch alone, measured on the first-phase model.

VariantGenEvalHPSv2.1
Both branches0.6425.42
Wide branch only0.6025.34
Codebook branch only0.5925.04

Joint training across tasks

Adding image-to-image and dense tasks costs little text-to-image quality, with DPG at 78.24 against 78.61. Both runs receive about the same number of text-to-image updates.

Training tasksGenEvalHPSv2.1
All tasks0.7428.06
Text-to-image only0.7328.28

Each card changes one component. Numbers are comparable within a card, not across cards.

BibTeX

@article{liu2026jeveration,
  title   = {Jeveration: System One Models Now Generate},
  author  = {Liu, Yuhe and Li, Qi and Zhang, Shihua and Wang, Xinchao},
  journal = {arXiv preprint arXiv:2611.24738},
  year    = {2026}
}