Jeveration: a multi-task model for image generation and analysis in a single pass
2 min readupdated
Jeveration is a frozen vision-language model (a multimodal model that understands both text and images) that generates 512x512 pixel images as a unified solution in a single forward pass. Built on top of the Qwen/Qwen3.5-4B backbone, it handles a wide range of tasks—from text-to-image generation and style transfer to depth estimation and segmentation. The model suits developers and researchers interested in all-in-one universal pipelines powered by modern language architectures.
Key features
- Single pipeline for everything: supports text-to-image, image-to-image (including HED/depth map processing, colorization, and styles), and dense prediction.
- Qwen3.5-4B base: uses a powerful language model as a foundation with added LoRA adapters (rank 256) and specialized heads.
- TA-TiTok decoder: employs a text-aware decoder fine-tuned for high-quality image output at 512x512 resolution.
- Dense prediction: capable of computing 13-class segmentation, depth, and surface normals for every pixel.
What you can do
- Text-to-image generation: create images from text prompts using a Python pipeline.
- Image transformation: convert photos into anime, ghibli, or watercolor styles, and colorize black-and-white snapshots.
- Structure processing: generate images based on edge maps (HED) or depth maps.
- Scene analysis: perform indoor segmentation and extract depth maps with surface normals using the dense prediction feature.
Example
from PIL import Image
from jeveration import JeverationPipeline
pipe = JeverationPipeline("weights/jeveration.pt", decoder="weights/decoder512.pt",
source_decoder="weights/source_decoder512.pt")
prompt = "A lone pine tree on a small island in a calm lake at dawn."
images = pipe.text_to_image([prompt], pipe.generator(0))
Image.fromarray(images[0]).save("pine.png")
Read more
Details• Model files: weights weigh about 4.0 GB, decoders are 1.2 GB and 1.3 GB.
• Output resolution: fixed at 512x512 pixels.
• License: Apache License 2.0 for code and weights.
• Limitations: dense prediction heads were trained on the NYUv2 dataset for interiors.










Comments