DreamOmni2 image edits were tested here as a reference-driven workflow, not as a prompt-only filter. Each test asks whether the model can read a source image, borrow a concrete object or a visual attribute from another image, and still make one convincing final image. The six examples cover wardrobe transfer, two-subject composition, pattern transfer, a three-image merge, style transfer, and a local repaint. Run DreamOmni2 on Wiro.
What the test set checks
The model card exposes two controls for these runs: 25 inference steps and guidance scale 3.5. Those are the settings used for every example below. The first image is the image to edit when the task is an edit; later images act as references. That ordering matters because the upstream instructions also place the editable image first. The Wiro documentation includes a one-output example that completed in 6 seconds. Its task record lists a total cost of 0.003510, so that is the only documented per-output cost quoted here. Actual elapsed time and cost can change with queueing and the selected run.
The useful question is not whether an output looks attractive in isolation. It is whether it respects the requested relationship between inputs: the person should remain the person, the car should remain the car, and a transferred pattern or style should not erase the scene. The official DreamOmni2 repository on Hugging Face, the project code on GitHub, and the research paper describe the same target: multimodal instruction-based editing and generation with concrete subjects and abstract attributes as references.
Settings and result notes
- Model: xiabs/DreamOmni2 on Wiro
- Inference steps: 25
- Guidance scale: 3.5
- Documented Wiro example: one PNG output, 6 seconds elapsed, total cost 0.003510
1) Outfit swap: identity first, clothing second
| Inputs | Output |
|---|---|
  |
 |
This is the cleanest practical editing case in the set. The output has to preserve the first subject’s face, pose, framing, and lighting while taking the outfit from the second image. It shows why a reference image beats a long text description for garments: cut, fabric, and small design cues travel together. Pick DreamOmni2 for a guided wardrobe concept when source identity and pose matter more than pixel-perfect catalog accuracy.
2) Two-character composition: a generation-style edit
| Inputs | Output |
|---|---|
  |
 |
This result asks for more than an edit. It creates a new scene, assigns left and right positions, and gives the characters an interaction. The output demonstrates that the references can serve as subject anchors while the prompt supplies a setting and action. This is the right mode when an existing photo does not need strict preservation and the goal is a new composition built around two known subjects.
3) Pattern transfer: the curved-surface check
| Inputs | Output |
|---|---|
  |
 |
The car makes this a tougher test than a flat product mockup. The transferred pattern needs to follow panels, curves, and reflections without turning into a pasted texture. The output is useful as a concept image because it clearly carries the second image’s visual motif onto the first image’s vehicle. For brand-pattern exploration, give the model a clean subject photo and a distinct reference. For production vehicle graphics, treat the result as art direction rather than final manufacturing artwork.
4) Three-image merge: subjects plus style
| Inputs | Output |
|---|---|
   |
 |
This is the hardest test in the article. DreamOmni2 must retain two different subjects, place them in a new environment, and borrow an abstract look from a third image. The result shows a coherent single frame instead of three unrelated references. Pick this approach for moodboards, character concepts, or story beats where subject identity and look need to arrive in one run. Use fewer references when strict source fidelity is the priority.
5) Style transfer: preserve the scene, change the visual language
| Inputs | Output |
|---|---|
  |
 |
The result tests whether the model can separate content from treatment. The output should still read as the first image, while palette, line work, texture, or atmosphere move toward the second. That is a better fit than a text-only style prompt when the reference has details that are hard to name. Choose this route for a fast art-direction pass. Use a single local edit instead if only one object needs changing.
6) Local car repaint: control the smallest change
| Before | After |
|---|---|
|
|
 |
The final output narrows the request to paint and a thin stripe. It is a good test of restraint: the scene and camera angle should remain stable while the object changes. This is the mode to pick for a single visible alteration, such as a colorway, material, or small product detail. State what must stay unchanged in the instruction, then inspect edges, reflections, and nearby objects before using the result.
When to choose DreamOmni2 image edits
Choose DreamOmni2 when the reference image carries the information that words cannot describe well: a specific outfit, texture, pattern, subject, lighting feel, or illustration style. Use one source and one reference for the tightest edit. Use two or three references when assembling a new image matters more than exact preservation. The six outputs make the tradeoff clear: the model is strongest as a visual-relationship tool, not a replacement for manual retouching where every pixel must match.
For another reference-led editing workflow, see Seedream V5 Lite image editing tests, FireRed Image Edit vs Seedream V5 Lite, and GPT Image 1.5 text and edit tests. Then try DreamOmni2 on Wiro with a clear base image, a purposeful reference, and a prompt that names both the intended change and the parts that must stay put.
