carlofkl/DreamLite-mobile

Requirements

This pipeline relies on Qwen3VLForConditionalGeneration / Qwen3VLProcessor. Due to upstream changes in transformers >= 5.0, you must pin:

pip install "transformers==4.57.3"                                                                                                                          

Using transformers >= 5.0 will produce visible block-pattern artifacts in the generated image.

DreamLite-Mobile

Distilled single-branch variant of DreamLite-base for fast, on-device inference. No CFG; runs in ~4–8 steps at 1024×1024.

import torch
from diffusers import DreamLiteMobilePipeline

pipe = DreamLiteMobilePipeline.from_pretrained(
    "carlofkl/DreamLite-mobile", torch_dtype=torch.bfloat16
).to("cuda")
image = pipe("a corgi astronaut", num_inference_steps=8).images[0]

License: CC BY-NC 4.0 (non-commercial). A full model card will be added once the diffusers integration PR is merged.