DreamOmni2 multi-image editing is the point of this six-case test: can one model keep a source image recognizable while borrowing a person, garment, pattern, style, or setting from one or more reference images? The outputs below are the original, unretouched results. This is not a pixel-matched benchmark. It is a practical check of subject retention, attribute transfer, and whether the new elements sit naturally in the scene.
What DreamOmni2 is built to do
DreamOmni2 on Wiro accepts an input-image list and a text instruction. Its documentation exposes two controls: steps (default 25) and guidance scale (default 3.5). The saved post records the input images and prompts, but not the historical task payloads, so it would be misleading to claim those defaults were confirmed for every output. The same applies to speed and price: the Wiro documentation includes a separate example task that completed in 6 seconds at a total cost of $0.00351; it is not a measured runtime or cost for any image below.
The ordering matters. The model maker’s editing instructions place the image to edit first, then reference images. That convention makes these prompts easier to read: image 1 supplies the scene or main subject; later images supply the garment, person, pattern, or visual treatment. For a closer look at the underlying model, see the DreamOmni2 Hugging Face repository, the official project page, and the paper.
Test rules and parameters
- Each case uses one output and one to three supplied images.
- Outputs are shown as generated, without manual retouching or compositing.
- Prompts name what should move and what should remain fixed where that distinction matters.
- The model documentation lists 25 steps and scale 3.5 as defaults, but the archived jobs do not preserve a per-case parameter record.
DreamOmni2 multi-image editing: six outputs
1. Outfit transfer
This checks whether the model can change clothing while retaining the person and location. The result shows a young man in a tan bomber jacket under the original curved glass walkway. The jacket has a centered zipper, ribbed cuffs, pockets, and believable folds. The face and hair remain coherent, though the output is a fresh render rather than proof of exact identity preservation.
2. Two-person scene composition
This is the harder identity test: two references must become one plausible interaction. The output places a woman in a striped blouse and a man with glasses in a sunlit terminal, with their hands meeting at the center. Faces, lighting, and the shared ground plane look consistent enough for a lifestyle illustration. The man’s shirt carries garbled logo-like text, so this is not the choice for branded apparel or readable type.
3. Pattern transfer onto a car
The pattern test asks for an abstract visual attribute, not a literal object swap. The car keeps a low sports-car silhouette and reflective panels while receiving a bright black, purple, and orange firework-like wrap. The result handles the difficult parts well: motifs bend around the doors, wheel arches, and hood instead of sitting as a flat rectangle. It does not reproduce a simple black-and-white mouse pattern literally, so it is better read as a creative interpretation than production artwork.
4. Cat and dog in one scene
This case tests two animal references and contact with a new environment. The output produces a cat and a corgi seated side by side on a dark car seat. Fur edges, ears, paws, and seat contact read cleanly at this size. The animals look polished, but the scene is more portrait-like than candid; review paws, whiskers, and breed markings before using an output for a client-facing asset.
5. Style transfer
The fifth output is the clearest result for a broad visual treatment. It converts the source into a soft colored-pencil beach scene with visible hatch marks, pastel sky, sun, water, sand, and birds. It clearly follows the request for brush texture and a new style. It is less useful when composition must be held exactly: the output reads as a new illustration, not a strict restyle of a photographed layout.
6. Background replacement
The final case isolates scene replacement. The orange sports car remains the subject while the daylight alley becomes a dark, rainy street. Rain streaks, wet pavement, red reflections, and low exposure make the car read as part of the new setting. The new image shifts the car paint from the source, so use this mode for mood and campaign concepts, not archival product photography where color must match exactly.
When to pick DreamOmni2
Pick DreamOmni2 when the reference image contains the thing text cannot describe well: a garment cut, a texture, a visual style, an animal, or a second subject. Keep the edit narrow when source fidelity matters. Name the source image first, say exactly what to borrow from each reference, and state the parts that must stay fixed. Use a different workflow when exact logos, typography, precise colors, or legally sensitive identity matching are non-negotiable. The six outputs make the trade-off plain: it is strongest at plausible multi-reference synthesis, not guaranteed forensic preservation.
Related image-editing tests
- FLUX Kontext Max Multi: 6 Two-Image Edits
- FireRed Image Edit 1.1: 6 Real Before and After Edits
- Top 5 Image Edit APIs for Product Photos
Try DreamOmni2 on Wiro
Run DreamOmni2 with the image being edited first, then add the references that supply the subject or attribute you want to carry across.
