Español · LocalText2Voice · YouTube channel · Report an issue
Video Storyboard turns an audiobook's narration into an editable visual timeline. It helps you plan scenes, keep reusable character and location descriptions, generate or bring your own images, create video clips, and assemble an MP4 with the audiobook audio.
It is a beta: useful for making narrated stories, illustrated lessons and video presentations, with a review step before export. AI can still confuse characters, invent details or place a scene at the wrong moment. The editor lets you correct the plan and individual results rather than regenerate everything.
This guide describes the Video Storyboard beta included in LocalText2Voice 2.0.1. Features marked beta may still change as the workflow matures. Updated: 13 September 2026.
- Analyze text to propose scenes and their opening sentences.
- Build editable character and location profiles, including visual descriptions.
- In full analysis, look for explicit appearance changes and create timed states.
- Review the initial reports before continuing, when review is enabled.
- Use a manual historical setting or detect one/multiple periods automatically, review their visual context, and assign them to scenes. See Historical periods.
- Set a maximum image duration; longer proposed scenes are divided into shots.
- Inspect the narration alongside the images and adjust scene boundaries.
The analysis uses narration timing cues when available. Without cues, timing is estimated from text. Even with cues, an opening sentence can be aligned to the start of its narration block rather than to its exact spoken word.
For long audiobooks, each proposal retains its source passage. Quotes are checked only against that passage; global era reports do not supply scene content. When TTS preparation changes the original text, analysis uses the spoken text and saves that coordinate reference. Failed anchors receive bounded local repair and replanning; a remaining failure stops analysis and identifies the passage to review. Unanchored scenes are never interpolated between distant points in the book. The Analysis checks tab shows validation results and warnings; ≈ beside a scene identifies estimated sentence timing.
Repeated dialogue and refrains use preceding and following quotes to identify a unique position where possible. Remaining ambiguity requests a longer literal quotation. Missing titles are filled without changing narration timings. After a failure, Resume saved analysis and reports reuses complete reports, finished entity profiles and saved proposals. Saved scene image prompts are also reused, continuing with the next missing prompt after credit or connection errors. Quotes, narration and timings are revalidated; text and settings must match. Saved progress is preserved if that verification is interrupted or finds an incompatibility. Uncheck this option to start a new analysis.
Under Settings → Video Storyboard → LLM, Retries on temporary errors sets additional attempts per request (default 5; 0 disables retries). Temporary connection failures, timeouts and service overload retry only the affected request, with cancellable delays increasing to 30 seconds. Authentication, credit and response-format errors are not retried by this mechanism.
Analyze audiobook → Import / edit scene prompts bypasses in-app LLM analysis. The first tab provides narrative, educational stickman and documentary request examples. Copy one with the project's timed segment text, or load a final-audio transcript (SRT, VTT or TXT). Times must use the final storyboard timeline, including any opening voice delay; segment timings are not word alignment.
Paste or load a TXT in the second tab:
[00:00] A red car drives along a New York avenue.
[00:06.500] Close-up of the driver looking through the window.
Each image lasts until the next timestamp; the last ends with the audio. Hours, decimal commas and multiline prompts are supported. Preview validates ordering, duplicates, empty prompts, audio bounds and the renderer's minimum scene duration of 0.1 seconds. A later first timestamp requires explicitly choosing to show that image from the beginning. Validation checks format and timing, not factual accuracy or the visual interpretation.
Use Complete prompts for self-contained descriptions: project style, narrative context, exclusions and automatic references are not added; manually selected image references still apply. Turn it off to use the project's style.
Reopening the dialog loads all current prompts. Applying replaces the timeline without generating images or calling an LLM. Unchanged scenes retain their media; changed scenes get new IDs to avoid stale generated results. Files are never deleted by import. The operation supports undo/redo and normal project saving.
- Generate storyboard images and regenerate individual results.
- Choose a visual style, resolution and project-specific visual settings.
- Edit a scene's image prompt and inspect the effective generation prompt.
- Edit characters, locations and their appearance states.
- Import or paste an image, fit it to the frame, or copy an existing frame.
- Use a configured AI image editor to revise a picture with instructions.
- Add character or location reference images for supported reference workflows.
The scene description and the reusable appearance profiles are combined when preparing the image-generation prompt. This encourages continuity; it does not guarantee identical faces, clothing or geometry in independently generated images. Reference-image support depends on the selected provider and workflow.
- Generate a video clip from a scene image, individually or in a batch.
- Preview clips and regenerate them with revised motion instructions.
- Import video through the storyboard's media workflow.
- Trim a clip or remove a selected section in the video editor, with undo/redo.
- Copy a clip's last frame to help prepare the next shot.
- Split or remove image frames and adjust timing in the timeline.
- Combine still images and video clips, with configurable motion and transitions.
- Render a final MP4 using the audiobook audio and open the output folder.
Clip lengths and reference-frame options depend on the video model. This is a narration-focused editor, not a promise of automatic lip sync or a replacement for a general-purpose multitrack video editor.
Configure the services in Settings → Video Storyboard (Beta). Text analysis, image generation, image editing and video generation are separate tasks.
| Task | Supported routes in the current UI |
|---|---|
| Text analysis | Ollama or a LiteLLM/OpenAI-compatible endpoint |
| Images | ComfyUI Z-Image-Turbo, a custom ComfyUI workflow, compatible image APIs through the LiteLLM option, or Runpod |
| Image editing | Configured ComfyUI editing workflows, compatible image-editing APIs, or Runpod |
| Video clips | ComfyUI workflows, including the available Wan/LTX profiles and custom workflows, or Runpod |
| Final assembly | Local FFmpeg rendering |
Available models and controls vary by route. Selecting a provider does not mean that every model supports every feature. Local services need their models and workflows installed; GPU memory requirements vary substantially, especially for video. Remote services require their own credentials and may charge for use. With local services, processing can stay on your machine. Remote services receive the text and/or media needed for the requested operation.
The basic image settings keep resolution and style accessible. The former advanced image card has been removed; workflow-level customization belongs in ComfyUI. Existing saved sampling parameters remain compatible with the app.
- Prepare a short narration and generate its audio in LocalText2Voice.
- Configure the analysis model and image provider in Settings.
- Open Video Storyboard (Beta) and analyze the audiobook. Enable report review if you want to inspect the extracted information before scene generation.
- Check names, appearances, places and scene timing. Correct errors early.
- Generate images, review them against the narration and replace or regenerate the ones that need work.
- Optionally animate selected images. Review and trim the resulting clips.
- Preview the sequence, check transitions and render the final video.
Start with a few scenes to learn your model's behavior and generation cost before processing a long book. Project data stores the scene plan, media references and continuity information so you can return to the work.
- Identity: an analysis model can merge names, miss people or invent them.
- Appearance: unspecified hair and clothing can be chosen as an illustrative design. They are not verified biographical details, including in true crime.
- Narrative continuity: the model may show the wrong person, anticipate an event or confuse a flashback. There is no complete automatic factual audit.
- Places: scene proposals feed the location summary, so invented details can propagate into a location profile.
- Timing: repairs and approximate alignment can leave an image early or late.
- Generated media: faces, hands, clothing and motion can vary between shots; provider failures, model availability and memory limits also affect results.
Check the complete sequence before publishing. Fixing one prompt can improve a particular example without establishing that the problem is solved for all books.
The focus is to make each stage understandable and editable: discover the cast, describe appearances, propose scenes, align them to narration, then generate media. Shorter, focused model requests make failures easier to inspect.
One test identified several people correctly in the initial report, but the
conversion step named every JSON entry Human. The application interpreted
those repeated names as one character and retained the first portrait. Later
steps inherited the broken cast.
We tested shorter instructions that explicitly separate a person's name from their species and require hair and clothing in the visual description. A recent test preserved all six named people and supplied hair and clothing for each. Some descriptions still omitted hair length. This is progress on that test, not a guarantee of character continuity or a completed identity-validation fix.
Other recent work includes reviewable analysis reports, recovery from empty image-description replies, adjustable maximum shot duration, timeline editing, image fitting and video trimming. The next priorities are stronger identity and narrative checks, more reliable timing, and clearer recovery when providers fail. These are development priorities, not promised delivery dates.
For implementation details, see Storyboard continuity. That document contains dated notes, including older approaches retained for context.
The official channel is LocalText2Voice on YouTube. Watch 50 AI Image Styles for Video Storyboards | VideoStoryboard Demo for a showcase of the built-in visual styles. The app includes 50 style presets with preview images, plus a custom style prompt. More demos and walkthroughs will be added to the channel as development progresses.
To report a problem, open a GitHub issue with the app version, analysis model/provider, affected step, expected result and actual result. A short reproducible passage and screenshot are especially useful. If you attach analysis logs, review them first: they can contain the narration and generated prompts. Remove private text and credentials before sharing.