mirror of
https://github.com/hanzoui/workflow-templates-broken-zap-2026-06-08.git
synced 2026-10-30 20:44:35 +00:00
1.9 KiB
1.9 KiB
OmniGen2
OmniGen2 is a multimodal generation model with dual decoding pathways for text and image, built on the Qwen-VL-2.5 foundation by VectorSpaceLab.
Model Variants
OmniGen2
- 3B vision-language encoder (Qwen-VL-2.5) + 4B image decoder
- Dual decoding with unshared parameters for text and image
- Decoupled image tokenizer
- Apache 2.0 license
OmniGen v1
- Earlier single-pathway architecture
- Fewer capabilities than OmniGen2
- Superseded by OmniGen2
Key Features
- Text-to-image generation with high fidelity and aesthetics
- Instruction-guided image editing (state-of-the-art among open-source models)
- In-context generation combining multiple reference inputs (humans, objects, scenes)
- Visual understanding inherited from Qwen-VL-2.5
- CPU offload support reduces VRAM usage by nearly 50%
- Sequential CPU offload available for under 3GB VRAM (slower inference)
- Supports negative prompts and configurable guidance scales
Hardware Requirements
- Minimum: NVIDIA RTX 3090 or equivalent (~17GB VRAM)
- With CPU offload: ~9GB VRAM
- With sequential CPU offload: under 3GB VRAM (significantly slower)
- Flash Attention optional but recommended for best performance
- CUDA 12.4+ recommended
- Default output resolution: 1024x1024
Common Use Cases
- Text-to-image generation
- Instruction-based photo editing
- Subject-driven image generation from reference photos
- Multi-image composition and in-context editing
Key Parameters
- text_guidance_scale: Controls adherence to text prompt (CFG)
- image_guidance_scale: Controls similarity to reference image (1.2-2.0 for editing, 2.5-3.0 for in-context)
- num_inference_step: Diffusion steps (default 50)
- max_pixels: Maximum total pixel count for input images (default 1024x1024)
- negative_prompt: Text describing undesired qualities (e.g., "blurry, low quality, watermark")
- scheduler: ODE solver choice (euler or dpmsolver++)