How to generate images and videos in ScreenWeaver: references, models, durations and drafts

Working by hand is perfect for a single shot you want to control. The professional order: a strong still first, then motion. Fixing a still costs seconds, fixing a video costs a full generation, and a good frame becomes your start frame.

Click the Image tile in the toolbar, then click inside the scene's Production group. The new node takes the project aspect ratio, since that setting applies to newly created shots and image-generation nodes.

123456789

Swipe sideways to see the whole screen.

1. 1References. The entity sheets and art style wired into the node (@characters, @places, @art_direction). Wire a reference in, then cite it in the prompt by its @handle.
2. 2Image preview. The current image of the node. Its output can feed a video node as a start frame or a reference.
3. 3Takes. Every image this node has generated. Browse them with the arrows and save one with the download icon; running again does not erase earlier takes.
4. 4Show what the model receives. "Show what the model receives" switches to the final prompt sent to the model; "Back to editable prompt" returns to your text. Check it when a result ignores an instruction.
5. 5Prompt. Your instruction, with @references to the wired sheets. Spend the words on framing, action and light; the look comes from the art style reference.
6. 6Model. GPT Image 2.5 Sunburst (default), Nano Banana 2.1, Seedream 5.0 Pro, GPT Image 2.5 Flare or Grok Imagine Image.
7. 7Quality. Low, Medium, High, Extra high or Max. Explore at a lower quality, render the chosen frame higher, and check the cost in the Generate tooltip.
8. 8Generate. Runs the node. The credit cost shows on the button and in its tooltip, and the job appears in Generations.
9. 9Images per generation. 1, 4 or 8 images per run. Generate several to choose a composition, one once you know what you want.

1. 1Plug in the references (marker 1): the production references of the characters, place and objects in the shot, plus the Art Style Image. They arrive as purple image wires.
2. 2Write the prompt (marker 5). Type @ to call each reference by its handle instead of describing it again.
3. 3Choose the model (marker 6): "GPT Image 2.5 Sunburst" is the default, with "Nano Banana 2.1", "Seedream 5.0 Pro", "GPT Image 2.5 Flare" and "Grok Imagine Image" in the same menu.
4. 4Set "Quality" (marker 7): Low, Medium, High, Extra high or Max.
5. 5Set "Images per generation" (marker 9): 1, 4 or 8.
6. 6Click Generate (marker 8). The button shows the credit cost before you spend anything.

Every run lands in the takes strip under the preview (marker 3). Keep the take where faces, costumes and light match your sheets. You can also upload your own image instead.

Add a Video node and you will not find a mode switch. The node reads what you plug into it.

1234

Swipe sideways to see the whole screen.

1. 1Wired image. A production image connected to the video node. As a start frame it fixes the first frame, so the clip keeps your composition, faces and light.
2. 2Input ports. The ports on the left of the node: "Start frame", "End frame", reference images and text. "Reference images are ignored while a start frame is connected."
3. 3Video node. The node that generates the clip ("No video" until a run). Its prompt, model, resolution, duration and audio controls sit below the preview.
4. 4Four ways to drive a video. Prompt only (text to video), Start frame (image to video), Start + end (between two frames) or Reference images (reference to video). Models differ: Kling v3.0 i2v needs a start frame, and Veo 3.1 reference-to-video is 8 s only.

• Prompt only (text to video): words only. Maximum freedom, minimum consistency.
• Start frame (image to video): an image on the "Start frame" handle becomes the first frame of the clip.
• Start and end (between two frames): add an image on the "End frame" handle and the model finds the way between them.
• Reference images (reference to video): images that define who and what appears, while the model stages the shot itself.

123456789

1. 1Video node. The clip preview ("No video" until a run). Its "Generation history" strip keeps your takes and versions, and "Edit video" works on a result.
2. 2Inputs. The ports for the start frame, the end frame, reference images and text. Reference images are ignored while a start frame is connected.
3. 3Show what the model receives. Shows the final prompt sent to the model; "Back to editable prompt" returns to your text.
4. 4Prompt. Describe the motion, the camera move and the sound, citing characters and places with @references. A counter ("{current} / {max}") shows how much room the model allows.
5. 5Model. Seedance 2.5 is the default. Each setting first shows its common values; the rest appear under "All options".
6. 6Video models. Seedance 2.5 (recommended), Veo 3.1 (16:9 and 9:16 only), Seedance 2.0, MiniMax H3, Kling v3.0 i2v (needs a start frame), Kling v3.0 t2v, Kling v3.0 Motion Control (may be blocked without video references) and Grok Imagine Video 1.5. The Kling models add a Mode, Standard or Pro.
7. 7Resolution. 720p or 1080p on most models, 2k on MiniMax H3, 1080p and 4k on Veo 3.1 at 8 s only. Seedance 2.5 adds a 480p "Draft" ("Low-resolution draft for a quick, cheaper preview.") that you can finish within 7 days with "Finish seedance draft".
8. 8Duration. Seedance 2.5 and 2.0: 4, 5, 6, 8, 10, 12 or 15 s (Seedance 2.5 goes up to 30 s). Veo 3.1: 4, 6 or 8 s. Kling v3.0: 5 or 10 s. Grok Imagine Video 1.5: 4, 6, 8 or 10 s. MiniMax H3: 5 to 15 s. Match it to the shot.
9. 9Audio on / off. "Audio on" or "Audio off". MiniMax H3 and Grok Imagine Video 1.5 always generate audio; Kling v3.0 Motion Control has none.

One row holds the settings: model, resolution, duration and "Audio on" / "Audio off", then Generate with its cost. A counter under the prompt shows its length against the model's maximum. The values listed first are the common ones; the rest sit under "All options".

• Seedance 2.5 (default): 4, 5, 6, 8, 10, 12 or 15 s shown first, up to 30 s; 720p or 1080p, plus a 480p Draft.
• Veo 3.1: 4, 6 or 8 s; 720p or 1080p. 1080p and 4k are 8 s only, reference to video is 8 s only, and only 16:9 and 9:16 are supported.
• Seedance 2.0: 4, 5, 6, 8, 10, 12 or 15 s; 720p or 1080p.
• MiniMax H3: 5 to 15 s; 2k; audio always on.
• Kling v3.0 i2v: 5 or 10 s; 720p or 1080p; needs a start frame; Mode Standard or Pro.
• Kling v3.0 t2v: 5 or 10 s; 720p or 1080p; Mode Standard or Pro.
• Kling v3.0 Motion Control: 5 or 10 s; 720p or 1080p; no audio; may be blocked without video references.
• Grok Imagine Video 1.5: 4, 6, 8 or 10 s; 720p or 1080p; audio always on.

Seedance 2.5 is the recommendation for narrative scenes, but there is no best model, only the best model for a shot. Test one shot with the same references and prompt on two or three models, and score the results.

In the Seedance 2.5 resolution menu, "480p · Draft" is a "Low-resolution draft for a quick, cheaper preview." Use it to test motion and performance. Generate your takes as drafts, pick the winner, then use "Finish seedance draft" to get it at 1080p. Drafts can be finished for 7 days, so finish your keepers the same week.

The Upscale tool under "Video tools" can do the same with "Seedance 2.5 Draft finish", and later take a finished video to 1080p or 4k with Topaz. That step is covered in finish and deliver.

A model fills every silence with a default. Write each shot in the same order: camera, subjects, action, performance, world and light, then sound and constraints.

• Always name a camera movement and its speed, or the model picks one for you: that is the floating camera.
• Write the action at the speed of the shot: five seconds holds two or three small actions.
• Show emotion as behaviour the camera can film, never as a feeling.
• For several shots in one clip, write a timeline (0 to 5 seconds, then 5 to 9 seconds) and say hard cut between shots.

Decide the sound shot by shot. "Audio on" can deliver dialogue and ambience in the take; if you will design the sound yourself, choose "Audio off".

The Video node keeps a "Generation history" strip, so a second try never erases the first. When you regenerate, change one thing at a time; change three and you will never know what fixed the shot.