Choose a Seedance model inside Dreamina by first deciding what your inputs need to control. An approved starting composition calls for a first-frame route. A specified destination calls for supported first and last frame inputs. References that separately guide appearance, action, camera or sound call for an appropriate multimodal reference route. Then check which model and controls your account actually offers.
A newer model name, a visible menu item and an available generation are different things. The useful choice is the configuration that supports your materials, required output and revision plan—not the largest feature list.
- Separate four decisions that often get mixed together
- Start with the materials you already have
- Use first-frame control when the opening composition matters
- Use Omni Reference when assets need different roles
- Read documented limits as mode-specific conditions
- Check account access before comparing candidates
- What historical configurations actually establish
- Do not merge tutorial labels into an invented interface
- Make the first generation answer a specific question
Separate four decisions that often get mixed together
Model, mode, account access and interface version answer different questions. Keep them separate while choosing a route.
For example, selecting Seedance 2.5 does not automatically make every uploaded image an exact first frame. Choosing a first-and-last-frame mode does not prove both slots contain images. A free account can display a model that requires an upgrade for the selected operation.
Record the complete configuration when comparing outputs: date, account, model version, mode, references actually supplied, duration, framing and quality settings. Without that record, a change attributed to “the model” may actually come from a different input role or unavailable setting.
Conceptual selection diagram, not a product screenshot or cross-model benchmark.
Start with the materials you already have
Choose the control you need before opening a long list of models.
Only a written idea: a text-based route lets the model establish the scene. You must describe its subject, action and setting because no supplied image fixes them. It suits exploration when the exact visual appearance is still open.
One approved image: a first-frame route gives the animation a visible starting composition. Use the prompt to describe the event, camera behavior and important anchors. A general appearance reference is useful too, but it is a different role from specifying where the clip begins.
An opening image and a target ending: use a mode that explicitly assigns those endpoint roles. The first-and-last-frame planning guide covers whether the images can plausibly connect. Two images in a generic reference area do not automatically define the endpoints.
Several assets with different jobs: consider a supported multimodal reference route. One image may guide a character, another the environment, and a video the desired movement. The task is to assign their roles and resolve conflicts, rather than upload every available file.
An existing video to change or extend: inspect the relevant editing or extension entry and its source requirements. Generating a fresh video from an image is not the same operation as modifying an interval in the existing clip.
The text-to-video versus image-to-video decision develops the first two routes under a common scene brief. Here, the next step is checking whether the desired input role is actually available in the selected Dreamina configuration.
Use first-frame control when the opening composition matters
A first-frame route suits an approved poster, product photograph, landscape or original illustration whose existing arrangement should begin the shot. The image supplies visible appearance and placement; the prompt supplies what happens next.
Before submitting, inspect the actual starting-image thumbnail. Check the correct file, crop and version. If the mode offers a last-frame input and you leave it empty, record that the run used only a starting image. Do not later describe it as a two-endpoint test.
The opening image does not lock every pixel through time. A camera move can crop information, a turn can reveal invented surfaces, and a generated action can redraw fine details. Choose a restrained motion when preservation matters, then review the whole result.
For a product whose exact label must stay visible, the selected input mode is only one decision. You may need to retain the real product on a separate layer or use suitable real footage. A prompt asking for unchanged packaging does not convert image-to-video into a guaranteed label-preservation tool.
First-frame control is also different from a reference image that guides a new shot’s appearance. If the exact opening arrangement matters, look for a control that assigns the image to that starting role. If it is absent, revise the route rather than assuming the general upload does the same job.
Use Omni Reference when assets need different roles
Dreamina’s current Seedance 2.5 usage guide describes AI Video > Seedance 2.5 > Omni reference, with + to add references. It then describes linking assets through @ tags, setting available video options, generating, reviewing and downloading.
Use the actual asset names or tags displayed in your workspace. State what each reference contributes. An image can guide appearance; a video can guide movement or camera rhythm; audio can guide the sound where supported. Supplying a file is not enough to resolve what should be copied from it.
Consider an illustrative project with an original blue delivery robot and a courtyard scene. A reference-role plan could be:
An untested example instruction might be:
Use the robot image for the character’s appearance and the courtyard image for the setting. Use the motion reference only for the gentle camera pace. The robot makes one small greeting gesture in a continuous shot. Retain its blue finish and distinctive markings; no dialogue.
Add the workspace’s actual asset tags when using the supported reference mode. This example is a proposed role assignment, not a completed multimodal test or proof of exact character retention.
Resolve incompatible references before generating. If one image shows a low camera and another a steep overhead view, decide which establishes the scene. If a motion video contains a rapid orbit but the brief requires a fixed viewpoint, choose a more suitable reference or change the plan.
More references do not automatically provide more useful control. Keep only the files that contribute something necessary and name their priority when they disagree.
Read documented limits as mode-specific conditions
Dreamina’s Seedance 2.5 product manual separates standard generation, documented at 4–30 seconds, from Long Video, documented at 30–180 seconds. The public tutorial linked above summarizes extended generation more broadly as 5–180 seconds; it does not provide the same mode-by-mode breakdown. Check the actual mode and available duration in your account rather than treating those descriptions as interchangeable settings. Extension has its own source and added-duration conditions.
The product manual also specifies a combined reference capacity of up to 50 inputs: up to 30 images, 10 videos and 10 audio segments, with additional length conditions on reference media. These detailed limits come from the manual, rather than the public tutorial. They are not a recommendation to fill every slot or evidence that every combination is equally stable. A small, purposeful reference set is easier to explain and review.
For base generation, the manual lists 480p or 720p output options. Keep native generation separate from later upscaling and the final editor export. A page headline mentioning 4K does not establish that the operation in your account generates a native 4K file.
The output workflow is MP4. If your delivery needs another format or a particular frame size, include the necessary finishing step instead of assuming the generation mode supplies it.
These are documented capabilities and conditions, not results measured across every account. The saved demonstrations discussed below used short clips. They do not validate long-video duration, maximum reference counts, exact timing, local preservation or every supported input combination.
Check account access before comparing candidates
Open the model selector and inspect the controls available with your account. A displayed option can still be gated for the chosen settings. Check the selected operation’s quote, allowance and any upgrade requirement before submitting.
If the desired configuration is unavailable, choose among three practical responses: use another available model and revise the brief accordingly, change the construction, or resolve the required access before proceeding. Do not describe a fallback as equivalent without checking the output it can deliver.
A lower-cost model may be suitable for composition exploration while failing a strict identity or information requirement. A higher-cost operation may still miss the brief. Compare against the same acceptance conditions and include all attributable attempts when evaluating cost. The usable-video cost guide explains that calculation.
Keep free and paid access observations dated. Access to one model on one account does not establish access to every model or the same allowance on another account. Promotional use, account credits and current subscription terms are separate conditions.
Before committing the project, use the limited-credit project worksheet to connect your configuration with the number of attempts and the required final export. Model selection is incomplete if you can generate a preview but cannot deliver the needed file.
What historical configurations actually establish
The following runs were recorded in a browser on October 14, 2026. The analysis uses saved settings, operation records and original outputs. These are dated examples of configuration differences, not a ranking or a current access guarantee.
The first product example downloaded as an 834 × 1112 MP4, with approximately 5.02 seconds of video. Its 500 ml marking was outside the frame at the reviewed 3.75-second checkpoint. The 720P selection therefore did not specify a literal 1280 × 720 file, and the input image did not establish continuous label visibility.
In the free-account example, Seedance 2.0 Mini was gated before a submitted job. The operation that actually ran used Seedance 1.0 Mini and consumed 25 credits. Its output was a 960 × 960 MP4 with visible watermarks and no audio stream. It was not the same model, quality selector or export as the paid product run.
A separate Basic-account product run with Seedance 2.0 recorded a 110-credit charge, compared with 210 credits for the earlier Seedance 2.5 product operation. Those records show different dated job charges under their settings. They do not establish that one model is generally cheaper, better or more faithful across tasks.
The detailed visual reviews used thumbnail sequences and native-size checkpoints, not full-resolution checking of every fine feature. Where audio streams existed in the paid runs, they were not heard and assessed. Keep those coverage boundaries separate from the configuration facts.
Do not merge tutorial labels into an invented interface
Dreamina’s published Seedance 2.0 tutorial describes a mode labeled Single-frame for first and last image inputs, and separately describes Multiframes for multimodal references. The Seedance 2.5 guide describes Omni reference. The saved short-run interface used first-and-last-frame wording.
Those names belong to particular documented or observed routes. They are not interchangeable instructions for every current screen. Follow the model and interface available to you, verify the role of each input, and retain the exact displayed names in your run record.
The Seed team’s Seedance 2.0 model overview documents multimodal inputs at the model level. Model-level support does not tell you which controls every app or account exposes. Dreamina’s product interface and the selected operation determine the actionable route inside Dreamina.
Likewise, a reference workflow described by another hosting service is not a source for Dreamina button names, credit rates or account permissions. Use it for a conceptual distinction only after checking Dreamina’s actual documentation and controls.
Make the first generation answer a specific question
Once a suitable configuration is available, use a representative short shot that tests the requirement you care about. If identity matters, include the detail that makes the character recognizable. If product information matters, review the actual wording and framing. A generic scenic demo cannot establish either requirement.
Keep the first brief manageable. Record the input roles, one primary action, the camera behavior and the anchors to preserve. Confirm available duration, aspect ratio and quality settings, inspect the quote, then generate and download the result when ready.
Compare the output with the brief before adding more references or increasing scope. If it fails, identify whether the issue belongs to the source, input role, motion request, account limitation or required finishing. Change the relevant decision instead of switching model names without understanding what changed.
Also inspect the revision route you will need. A temporal segment selection is not necessarily a spatial mask; a regenerated candidate is not a localized edit. If exact preservation outside a correction matters, verify that capability and review the resulting clip rather than assuming it follows from the model’s feature description.
Finish by checking the actual file: dimensions, duration, visible information, movement and required sound. The best available Seedance configuration for your project is the one that can carry your supplied assets through a verifiable generation and delivery path.