When OmniGen2 was announced for ComfyUI, it sounded like a dream: one model that could handle text-to-image, natural language image editing, and even multi-image composition—all through clean, natural instructions like “make her smile” or “replace the dress with a blue one.” On paper, it reads like a direct competitor to Flux.1 Kontext, the current go-to model for in-context editing and image refinement.
According to Comfy’s official tutorial, OmniGen2 even leverages Qwen-VL vision capabilities under the hood, promising high-fidelity understanding of both image content and instructions. The node integration is simple enough: load the model, plug in your prompt or reference image, and let it do its thing.
But How Well Does It Actually Work? That’s where things got disappointing.
After running a series of test cases—mainly editing—OmniGen2’s results were inconsistent. Sometimes the instruction was ignored. Other times, the model degraded the image greatly.
Examples
Here are some examples compared to Flux.1 Kontext—I’ll let you be the judge.
Input image
Prompt:
Remove the girl from the picture.
OmniGen2
Flux.1 Kontext
Prompt:
Add a dog to the right of the girl.
OmniGen2
Flux.1 Kontext
Prompt:
Change the outfit color to white.
OmniGen2
Flux.1 Kontext
Prompt:
Generate a ghibli anime style of this picture
OmniGen2
Flux.1 Kontext
Input image:
Prompt:
Generate a realistic photo of this beautiful asian woman while preserving facial features, pointy chin, and race.
OmniGen2
Flux.1 Kontext
Conclusion
In these comparisons, I personally prefer the outputs from Flux.1 Kontext—except for the last example. Here are a few advantages I found Flux.1 Kontext has over OmniGen2:
-
Facial identity was better preserved
-
Edits followed the instructions more accurately
-
Overall output quality was higher
So, what do you think? Which model do you prefer?
P.S. If you’d still like to try OmniGen2 for yourself, the outputs have the workflow attached. Just download the full-size image and drag it onto your ComfyUI canvas to load it.
Further Reading
How to Use Flux.1 Kontext-dev in ComfyUI (GGUF + GPU): Natural Prompts for Any Style
Reference
Leave a Reply