What Agnes Video 2.5 Can Do: 8 Prompt Examples

👁 115 reads

Agnes Video 2.5 operates across three primary modes: text, keyframes, and reference inputs (image, audio, and video). Here are 8 concrete examples: one clip published by the creator, six official prompts drawn directly from documentation, and a demo of the older, open-weights 2.0-MoE model. For full pricing, benchmark rankings, and setup instructions, see my guide to Agnes Video 2.5.

Prompts 2 through 7 are sourced from the Agnes Video 2.5 documentation (checked October 11, 2026). The input media URLs in those examples are dummy example.com addresses, so neither source files nor render results are available to review.

1. Horse in a forest stream (Agnes Video 2.5 Preview)

Use case: Nature scenes, wildlife, liquid and water dynamics. This is the only clip from the 2.5 lineage that the developer published alongside a descriptive scene prompt.

A wild horse crosses a shallow forest stream in an autumn valley with realistic motion, water interaction, and environmental response.

Agnes Video 2.5 Preview enables cinematic 3D animation with coherent motion and physics-aware consistency.

- Agnes AI (@agnesai_sapiens) July 2, 2026

Context: This text is the description from the social media post, not necessarily the verbatim generation prompt used by the system. The Preview version predates the model evaluated on public leaderboards. The clip runs for 15.7 seconds, whereas the public API currently caps single-request renders at 12 seconds; it was either stitched in post-production or rendered with internal parameters. The developer has not clarified this discrepancy.

2. Rain-slicked street with neon lights

Use case: Atmospheric B-roll for music videos or title sequences. An official prompt without public sample outputs; documentation settings specify 5 seconds, 720p, 16:9.

A futuristic city street after rain, neon lights reflected on the pavement, a silver sports car passing slowly, cinematic camera movement, natural ambient sound

Context: The prompt adheres to the developer’s recommended structure: setting, action, camera movement, and audio. Generating this clip costs $0.125. The phrase “cinematic camera movement” remains generic; for practical production work, I would replace it with concrete directions, such as a slow push-in or a vehicle tracking shot.

3. Three cats playing musical instruments

Use case: Stylized storybook sequences requiring audio synced to visible action. An official prompt without public render samples, featured in the docs’ sample curl payload (5 seconds, 720p, 16:9, seed 1101).

At night, three cats march through a forest while playing tiny brass instruments. The camera tracks backward smoothly, moonlight passes through the leaves, with natural footsteps and instrument sounds.

Context: This serves as a solid audio benchmark. The model must align footsteps and musical output with physical movement. Specifying a fixed seed enables reproducible testing while tweaking individual prompt parameters.

4. Character turning toward a window (keyframe mode)

Use case: Continuity transitions between two planned shots, such as bridging storyboard frames. Official prompt without public outputs. Requires both start and end frame assets (first_frame, last_frame).

The person turns naturally from the first-frame pose and walks toward the window. Keep clothing and hair motion realistic, slowly push the camera forward, and transition smoothly to the last-frame composition.

Context: This represents the standard image-to-video workflow in Agnes. Note that on the Artificial Analysis image-to-video leaderboard, the model ranks 25th (1031 Elo), indicating you should anticipate lower consistency here than in pure text-to-video mode.

5. Character running across a meadow (image reference mode)

Use case: Maintaining character identity across multiple shots, such as recurring branded mascot content. Official prompt without public sample outputs.

Use the character and art style in <Picture 1> as reference. The character runs naturally through a flower field, maintaining a consistent appearance, filmed from a low tracking angle.

Context: Images are supplied via an images array (up to 8 files) and referenced inline using <Picture N> tags. The first 5 reference images are included in the base price; starting from the sixth, each incurs an extra $0.005 fee.

6. Motion and cuts synced to music (image and audio reference mode)

Use case: Short music clips or commercial promos timed to a pre-recorded track. Official prompt without public sample outputs.

Use <Picture 1> as the visual subject and design the actions and camera cuts around the rhythm of <Audio 1>, keeping the sequence natural and coherent.

Context: Audio input supports up to 3 files with a combined runtime between 2 and 12 seconds. Conditioning visual edits on custom audio is significantly more versatile than relying solely on generated sound, though without sample renders, timing accuracy remains unverified. The Flash tier also accepts audio references (up to 3 files).

7. Motion transfer into a new environment (video reference mode)

Use case: Transferring choreography or raw mobile footage into a completely different virtual environment. Official prompt without public sample outputs.

Follow the subject motion and camera rhythm of <Video 1>, changing the setting to a moonlit bedroom while preserving coherent timing.

Context: Input video files must run 2-12 seconds, weigh under 50 MB, and feature a framerate between 24 and 60 fps. The documentation example specifies start_seconds alongside require_audio: false. The duration of the reference video is billed on top of the output at the chosen resolution tier. The Flash tier does not support video reference mode.

8. Cat and dog demos from Agnes-Video-2.0-MoE

Use case: Baseline comparison against the vendor’s previous open-weights architecture. The generation prompts match the file names in the official repository, and the README confirms they were synthesized from text prompts.

A cat is chasing a laser pointer, then it suddenly sits down and starts licking its paws.

A dog is playing in the yard, then it suddenly starts laying on the porch.

Context: Both prompts incorporate sequential instructions (“then it suddenly”), testing the engine’s ability to pivot action within a single continuous shot. The video files are minimal in size (roughly 340 KB and 380 KB), reflecting low output resolution. In the repository README, benchmark quality evaluations were recorded at 448 x 256. The repository also stores clips of a monkey with an apple and a rabbit by a flower pot; two additional links in the README (a sunrise and horses by a river) returned HTTP 404 errors as of October 11. Read my overview of Agnes-Video-2.0-MoE open weights for more technical details.

Unresolved questions from available samples

• How 2.5 handles dialogue rendering and lip synchronization. The documentation provides no lip-sync specifications or supported language lists.
• Whether non-English prompting yields comparable fidelity. All official prompt templates are provided exclusively in English.
• The practical performance gap between the Flash tier and the flagship model. Both share identical features aside from reference video support, but comparative benchmark outputs have not been published.

To compare different video models against identical scene prompts, visit the Video Arena. Agnes is currently omitted because it lacks OpenRouter availability. For budget-conscious alternatives, see my breakdown of the cheapest video generator in the top 10 and the broader overview of AI video generators.

Frequently Asked Questions

Are there official sample clips for Agnes Video 2.5?

As of October 11, 2026, only a single annotated preview clip exists on X (the horse crossing a stream). No 2.5 generation samples were found on Agnes AI’s YouTube channel.

Can you test these prompts directly?

Yes, these strings come straight from the API documentation. For keyframe and reference modes, you must supply your own valid image, audio, or video URLs.

How much does it cost to test all 6 official prompts?

At 720p for 5 seconds each: 6 x $0.125 = $0.75, plus the billed duration of the input video asset in Example 7. Testing on the free Flash tier costs $0, though it excludes reference video mode and has been subject to 503 gateway errors.