AI media harness for planning, generating, and editing locally
Scene types
Videos contain ordered Scenes, which contain Shots. The Video library supports folders, grid/list views and sequential card or popout playback. Click a Video to select its settings; double-click or choose Edit Scenes in the right panel or context menu to edit its scenes. Each scene remains separate in Timeline. Existing projects open with their scenes grouped in a default Video.
Video settings provide Look, aspect ratio and resolution overrides, each defaulting to the project setting. Scenes inherit their Video's settings; an explicit scene Look takes precedence. Choosing Project default restores inheritance for that Video setting.
Videos also provide a shared prompt with reference chips, background music and a reference video. Each scene receives the reference range at its cumulative start time: a 10-second first scene uses 0–10 seconds, and the next starts at 10 seconds. Imported/generated media uses the complete asset; reference entries use their saved trim range. Reordering or resizing scenes updates these ranges for the next generation. A source that ends too early blocks generation.
For a music video, select a complete song on an empty Video and choose Create scenes for song to divide its full duration into roughly 10-second scenes. Add visual direction and an image reference to the shared prompt, or select a reference video: SlopFab sound references need visual conditioning and References/Singularity weights. Music guides generated audio; add the original song to Timeline when you want its exact soundtrack. Shared media references support Prompt, Single Shot, Continue, First & Last Frame, Backdrop and Pose scenes; each scene must last at least two seconds.
The Scene panel saves a type on each scene. Long Shot is retired; use Bridge to connect videos. Saved Long Shot scenes retain their generation and playback behavior, while new or duplicated scenes use the available types.
Select a generated Take and open Generate's arrow to upscale its saved latents 1.5x, 2x or 4x into a new Take. Escape or clicking outside closes the menu. Upscaled Takes can be enlarged again; continuations and bridges always use their original latents.
- Prompt is the default for ordinary video scenes. It supports shot descriptions, speech, Look, sound, music and references cited in prompts, without start- or last-frame selectors. Stored frame anchors and continuation links are inactive.
- Single Shot keeps its main prompt in the Shot panel and an optional Ending prompt in the Scene panel, both with reference support. It generates an ending first, then anchors the full shot to it; a blank ending prompt uses the main prompt. Starting video accepts any generated scene. Its final latents anchor only the beginning; the middle starts from fresh noise and is guided toward the generated ending. The former Continuation strength is ignored. Use matching original latent resolutions, at least two seconds, and updated workers. Generate all handles dependency order.
- First & Last Frame adds optional start- and last-frame anchors. Prompts follow the MiniMax base guide: no frames uses T2VA, first only I2VA, last only L2VA, and both FL2VA. Ending anchors target the final shot at the generated duration, including frame alignment. Additional subject/audio/video references retain their full-reference sections. Older scenes without a type retain this type when they have frame selections; otherwise they use Prompt. Switching types retains saved frame selections for later use.
- Backdrop generates the described subject and action against a selected green, blue, black or white background (green by default). The compiled prompt requests a uniform color across all shots, with no scenery, floor, horizon, texture, gradients or backdrop shadows. Prompt references guide the subject; their original settings are replaced. Stored frame anchors and continuation links are inactive. Apply Chroma Key to the generated timeline clip and select the matching backdrop or Auto to remove it. Generation still produces ordinary video: transparency is applied by the effect. Choose a color absent from the subject and describe action that fits inside the frame.
- Continue uses another scene's saved latents. Choose the source scene, beginning or end (default: end), and overlap in frames (default: 22; increments of 17 at 24 fps). Generate the source first, or use Generate All to render dependencies in order. Scene length controls the new segment. The saved video keeps Slopfab's full joined video and audio; adjacent source and continuation clips use that shared decode in preview and export, while trims remain relative to each scene. Regenerate older Continue scenes to preserve both sides of the decoded boundary. Old Previous scene selections become Continue scenes linked to their original source; reordering scenes keeps that link.
- Animate uses an Animate generator, one driving video and one repainted scene frame. It does not send a text prompt.
- Pose uses a selected video for body pose, movement and timing. The Scene panel places the video selector above the shot prompts, where optional references can define the target subject and setting. It uses the video's saved clip range without its soundtrack, attached stills or refmods. Only the selected pose video and references cited in the prompts are sent; frame anchors and previous-scene continuation are inactive. Pose requires a nonempty prompt and reference-capable MiniMax weights (References or Singularity), with the existing limits of three videos, nine images and 15 seconds of reference video.
- Character Replace gives each shot a source video selector, a new character reference selector and an optional description identifying the character to replace. The character reference must contain an image. Blank target descriptions mean the main character.
- Extend takes a source video in the Scene panel and uses the shot descriptions as the continuation prompt.
- Bridge takes start and end video references in the Scene panel and uses the shot descriptions as the prompt for the connecting segment.
Character Replace sends only references selected by its shots, in project order. Frame anchors, continuation, old action text and style settings are inactive in this mode. Switching back retains the shot descriptions and character selections. The video reference's saved trim and soundtrack setting apply to generation.
The generated prompt follows the MiniMax H3 full-reference guide: six sections, a video-editing task prefix, independently numbered image/video/audio payloads, and explicit retention instructions. It replaces the target's appearance while retaining motion, timing, camera work, environment and other characters. Character images define the replacement identity rather than anchoring a frame. Enabled soundtracks are reused in the corresponding shots.
Generation requires both references for every shot and respects the existing limits of three video references, nine images and 15 seconds of total video reference material. The generator picker selects the active engine for the queue: incompatible scenes are excluded from Generate All. Changing a scene type keeps the active generator; choose a compatible one when needed. Character Replace requires reference-capable MiniMax weights, such as References or Singularity; known First/Last Frame weights are rejected.
Extend and Bridge require reference-capable weights and SlopFab API 1.14. The app decodes the selected reference ranges with WebCodecs. At generation time the runtime encodes the start video's last 22 frames into VAE latents at the project canvas. Bridge also encodes the end video's first 22 frames. These guides share the generated segment's spatial coordinates and sit immediately before/after it in the model's temporal coordinates. The encoded boundaries use the existing reference cache. No source latent files or FFmpeg decoding are required.
Each source range must contain at least 22 frames at 24 fps (0.917 seconds); ranges still share the 15-second video-reference budget. Start/end ordering is independent of library order. Frame anchors, previous-scene continuation and other references are inactive in these modes. Sound and speech are generated from the prompt; source soundtracks are not copied. Scene length controls only the new segment, which is the only video added to the timeline. Add the source clips on either side when arranging the complete sequence.