Dreamomni2
Open-source multimodal model for image editing and generation.Edit and generate images from text and reference images, open source.
Highlights
What you can do with it
- Edit product photos by applying new materials or styles
- Build and revise design drafts with combined text and image cues
- Retouch portraits such as makeup or hair while keeping identity
How to use it
- 1 Open the DreamOmni2 editor on the website
- 2 Upload the source image you want to edit
- 3 Add up to two reference images for traits or objects
- 4 Type natural-language instructions for the change you want
- 5 Run the edit, then preview and download the PNG result
Who it is for
Pros
- Combines text and image cues in one editing flow
- Open-source model with a local inference option
- Strong identity consistency and multi-image support
Limitations
- Hosted use is priced from about $4.90 up to $960, and local runs need a CUDA GPU with enough memory
- Local setup demands technical skill and suitable hardware
- Complex edits may need several tries to reach the look you want
What DreamOmni2 is
DreamOmni2 is a unified, open-source multimodal model for instruction-based image editing and generation. It is built on Flux Kontext and Qwen2.5 VL and accepts both text and image instructions, so you can point to a reference while you describe a change. The model is designed to handle abstract attributes such as material, texture, and style, as well as specific objects, and it aims to keep the identity of the main subject consistent while editing. The maker reports benchmark results above commercial models. Weights, training code, and datasets are published on GitHub and Hugging Face, and hosted pricing ranges from about $4.90 to $960.
What you can do with it
- Edit an image using text instructions plus reference images
- Change materials, textures, and styles while keeping the subject consistent
- Work with multiple input images without pixel confusion
- Run the model locally for full control over your data
Who it is for
Photographers and designers can refine product shots and portraits with precise, reference-guided edits. Researchers benefit from open weights, code, and datasets for experiments and comparison. Everyday users who want a guided browser editor can still get complex results without deep knowledge. Anyone comfortable with a local setup can run it on their own hardware for privacy and control.
What to watch out for
Hosted use is not free: pricing spans roughly $4.90 to $960 depending on usage, and these rates can change, so check current terms. Running locally avoids hosting fees but needs a CUDA-compatible GPU with enough video memory, plus the technical skill to set it up. Complex edits often need several attempts to match the look you want. As with any generative model, results can vary, so review outputs before using them in published or commercial work.
FAQ
Is DreamOmni2 open source?
Yes. The weights, training code, and datasets are available on GitHub and Hugging Face.
Can I run it on my own machine?
Yes. You can download the weights and run local inference on a CUDA-compatible GPU with enough video memory.
What inputs does it accept?
It takes source and reference images in JPG, PNG, or WebP, plus text instructions, and outputs PNG files.
Why it is free
What the official page says： 免费
Free limits, trial length and features are decided by the official site and can change at any time.
Product preview
Screenshot archived from the listing page, for preview only.
