Mtrini-Imagine-1.0-7B
Compiwer AI's first text-to-image release: a LoRA adapter for HiDream-O1-Image, plus a fused BF16 GGUF.
Mtrini-Imagine is Compiwer AI's experiment in open image generation. Version 1.0 was trained with DiffSynth-Studio as a rank-16 LoRA on top of HiDream-O1-Image-Dev, using 4,993 training examples.
The adapter is about 98 MB. A separate GGUF release fuses all 360 LoRA layer pairs into the base weights. The card notes that generation with the o1.c runtime has not yet been officially validated for the GGUF.
Specifications
- Status
- Experimental
- Base model
- HiDream-O1-Image-Dev (opens in new tab)
- Architecture
- HiDream-O1 / UiT
- Method
- LoRA, rank 16 / alpha 32
- Precision
- BF16
- Framework
- DiffSynth-Studio
- Training examples
- 4,993
- Languages
- en, ar
- Focus
- Text-to-image, Image-model research, Local inference
Downloads
Usage
import torch
from diffsynth import ModelConfig
from diffsynth.pipelines.hidream_o1_image import HiDreamO1ImagePipeline
pipe = HiDreamO1ImagePipeline.from_pretrained(
torch_dtype=torch.bfloat16,
device="cuda",
model_configs=[ModelConfig(path="PATH_TO_HIDREAM_MODEL")],
)
pipe.load_lora(pipe.dit, "Mtrini-Imagine-1.0-7B.safetensors", alpha=1.0)Limitations
- Text rendered inside images is often incorrect or misspelled.
- May produce odd object relationships, visual artifacts and inconsistent small details.
- GGUF generation via o1.c is not yet officially validated; the GGUF file itself was verified after conversion.
- Licensed Apache-2.0 subject to the HiDream-O1-Image base model's license — review it before commercial use.