Scene inspector

Storyboard · Русский · Contents

What happens

The scene in your words. If you change it, «Update prompt from text» appears: the Writer rewrites the English to match your text and leaves your text as it is.

Camera

Static, push in, pull out, side tracking, camera backing away in front of the hero, following from behind, orbit, pan, rise. Calm movements are more reliable; a static camera is gentlest on the face.

Shot

Wide, medium, close-up. A close-up shows the face better, a wide shot shows the place.

Start

A new scene (from text, its own place) or a continuation — from the last frame of the previous scene or of any earlier one («Continuation — from the last frame of scene N»). The latter is a cutaway: you showed something else and came back to the moment where scene N ended. Such a scene opens its own chain (with a change of shot before it), and face drift is counted by frames: «from the frame of scene 2» is the third continuation if scene 2 is itself the second. The card says «from frame N».

Scene model

Which model shoots the scene and its continuations: «same as the video» or its own. That way, in one video, story scenes with faces are shot by H3, landscapes and action by LTX 2.5 (three times faster), singing to a track by LTX 2.3. Frame sizes differ between models: in the video, a scene is fitted to the video's frame. Shooting goes model by model — each loads once. A continuation has no model of its own: it is shot by its chain's model, because it starts from that model's frame.

Scene length

5, 10 or 15 seconds — as much as the scene's model shoots: LTX up to 10 s, H3 up to 15 s, WAN only 5. A storyboard on LTX is written in 10 s scenes from the start: 30 s is 3 scenes, 15 s is 10 + 5. In such a scene the actions share all 10 seconds, and the lines can be twice as long — about 20 words. Length is part of the key: change it and the scene and the continuations after it are reshot, the scenes after it shift.

First frame

On a new scene. «Start from photo» — the scene starts from your picture, the scene's model brings it to life; «From hero photo» — from the picture on the Idea screen. Then be sure to press «Rewrite for photo»: the Writer looks at the picture and writes the scene from it.

✗ Worse: Uploaded a photo of a lavender field and shot right away. The text had a meadow under clouds — the model drifted from the photo to the text, and the face changed.

✓ Better: Uploaded the photo → «Rewrite for photo» → the scene starts in the same field, in the same pose.

Last frame

LTX only: «End on frame» — the scene ends on your picture. Describe in the prompt how it gets there: camera movement, action. A model that can't do this shoots the scene without it — that's written under the picture. The frame is part of the key: change it and the scene is reshot.

Who's in the scene

In «Several» mode: checkboxes for who is in the shot. Only the looks of the checked characters go into the prompt; otherwise the model draws everyone. A continuation without checks takes the cast of the scene that opened its chain.

Lines

Under everyone in the scene — what they say, in your language, and how: normally, quietly, in a whisper, shouting, with a smile, angrily… Empty — silent. Each line becomes its own timeline line: [1s-4s] Character 2 whispers, in English: "I'm sorry." — the speaker is named the same way as in the heroes' description («The heroine», «Character 2») so the lips move on the right face. The scene's actions and lines share its length: a line gets as much as its words need (about 2.5 words per second, at least 2 s), the actions get the rest. Voices are added to the sound, everything else is quieter under them. Lines are shown on the scene card too. The «12 / 10 words» counter above the lines is how many words everyone has and how many fit (about 10 per 5 s, 20 per 10 s); over that, the counter turns yellow. On WAN there's a warning above the lines, «… doesn't talk: lines won't be heard» — they need H3 or LTX.

✗ Worse: «You're late. I waited for you all evening and you didn't even call»

✓ Better: «You're late. I waited all evening»

A 5 s scene holds about 10 words for everyone together; longer, and the speech gets rushed or cut off, and the scene gets the «check» mark.

«Suggest lines» — the Writer proposes lines for those in the scene, from the story around it; yours are replaced, and you can edit them right away. Editing a line reshoots the scene and the continuations after it; scenes without lines don't change.

Action lines

1–3 actions in English, one per line; on H3 they go onto the timeline [0s-2s] [2s-4s] [4s-5s], LTX and WAN get them as sentences in order. Editing them pins the scene.

✗ Worse: She walks forward slowly.

✓ Better: She walks toward the camera, the wind lifting her hair.

Without a direction toward the camera the model often turns people to face us but moves them away from the camera — the result is walking backwards.

✗ Worse: She turns, then smiles and waves.

✓ Better: She turns to the camera.
She smiles.
She waves.

One action per line: the model spreads them over the seconds.

Prompt for the model

The collapsed «Prompt for the model, EN» block: what the scene's model will get. Code builds it from the scene's fields, differently for each model. H3 — by the formula: style, hero word for word, place, camera, timeline, sound and tail (taken apart line by line at the top of this page). LTX — one paragraph: style, place, who is there, the action from start to end, camera, sound with the lines. WAN — who, where, movement, camera, style, no sound; for a continuation and a scene from a photo, only movement and camera: the rest is already visible in the frame. For LTX and WAN the text is captioned «LTX 2.5 will get» (or another model): «on a new shoot; from the same fields, in its model's order». The storyboard stores the H3 prompt: switch back to H3 and what H3 shot is taken from the cache.

Own prompt

«Write the prompt myself» moves the built prompt into the «Prompt for the model, as is» field, where you can rewrite it completely. H3 gets it word for word, LTX and WAN without the H3 markup (the «How to write a prompt» section); the Writer no longer touches the scene. «Preset» inserts a sample. «Restore the Writer's prompt» builds it from the fields again.

Continue seamlessly from <Picture 1>.

Realistic live-action cinematic look, soft natural light, film grain.

Scene: the same rooftop at dusk.

Camera: medium shot, static camera.

Timeline:
[0s-2s] She leans on the railing and looks at the city.
[2s-5s] She turns her head toward the camera and smiles.

Audio: distant traffic, a light wind.

No text, subtitles, logos or watermarks.

The first line is needed only by a continuation or a scene from a photo: that's how H3 knows it must start from a ready frame (LTX and WAN don't get it). How to write your own prompt, the hero's speech and the camera words — the «How to write a prompt» section.

Takes

«Another take» shoots this scene again with a new seed — now, on its own (a separate job in the queue); earlier takes remain. The thumbnails are all the scene's takes: a click chooses which one goes into the video and into the continuations after it (their frame changes — they'll be reshot on «Shoot video»); × removes a take from the list, the file stays in the cache and comes back if that seed is chosen again. A continuation needs the shot frame of the scene it starts from to get a take. The card says «take 2/3». In the montage a clip can use another take without touching the storyboard.

Scene LoRA

«Same as the video» — the set from the shooting panel; «Own» — a set and strength for this scene only, from the LoRAs of the model that shoots it; «No LoRA» — the scene is shot without them. LoRA is part of the cache key: change it and this scene is shot (and the continuations after it), the rest come from the cache. The card says «own LoRA» or «no LoRA».

Pin, Delete scene

The lock and deletion; the other scenes' times are recalculated.