OmniGen2 Natural Language Image Editing
Give OmniGen2 one or two concept images and a written instruction to generate or edit an image. It is an experimental multimodal model that is useful for combining characters and concepts.
- L40S
- Image editing
- Text to image
- Free template
How it works
The workflow loads omnigen2_fp16.safetensors with UNETLoader, the qwen_2.5_vl_fp16 text encoder with CLIPLoader, and ae.safetensors as the VAE.
Two images, Concept1 and concept2, are resized with ImageScaleToTotalPixels. Backgrounds can be removed with BiRefNetRMBG, and VAEEncode converts them to latents. Several ReferenceLatent nodes attach those latents to the positive and negative conditioning so the model can see both concepts.
Sampling uses SamplerCustomAdvanced with DualCFGGuider, BasicScheduler and RandomNoise. The widen_difference slider controls the guidance difference between the two guiders. Steps default to 50. A ttN hiresfixScale node and the 4x_NMKD-Siax_200k upscaler are included for a second pass. An optional RepeatLatentBatch makes several images per run, and image comparers show before and after.
The notes record a tested sampler setting of dpm_adaptive with kl_optimal over 18 steps at about 1.9 CFG.
Tips
- Reference the images by order in your prompt, such as the girl in the first image and the man in the second.
- Use the negative prompt to reduce deformed faces and blur.
- Lower the steps from 50 for faster drafts.
- Set the number of images above 1 to compare variations.
What you can use it for
- Combine two characters into one scene
- Wedding or portrait figures from two source photos
- Edit an image with a text instruction
- Experiment with a multimodal model
Run it in three steps
Open the template
Click Run on ComfyDeploy. The workflow, its custom nodes and models are set up for you on a cloud machine.
Add your inputs
Provide concept image 1 and concept image 2 in the playground or the ComfyUI editor.
Run, share or deploy
Run it on an L40S, share it with your team as a playground, or deploy it as an API endpoint.
Frequently asked questions
What GPU does this OmniGen2 workflow need?
It is set to run on an L40S GPU on ComfyDeploy. It loads omnigen2_fp16 and qwen_2.5_vl_fp16.
How good is OmniGen2?
The author describes it as an experimental multimodal model without major pretensions that can still produce interesting results in some cases.
How do I run it as an API?
Deploy it on ComfyDeploy and send the images and prompt to the endpoint. The external input nodes become the API fields.
