Maestro v.1.7.0 - BIG UPDATE! Long multi-clip videos in FL & REF models

MiniMax H3 native multi-window generation

Added native multi-window continuation to both First / Last and Omni, carrying recent motion and synchronized stereo audio into each following window for smoother transitions.
Added shared Multi-window controls with total duration, editable per-window prompts, optional AI planning, and independent hard-cut sequences.
Improved H3 sequence planning so actions, dialogue, camera cuts, sound, and story events advance across windows instead of repeating or finishing in the first clip.
Added exact-duration assembly, model-aware overlap handling, and saved runtime prompts so long generations can be reviewed, edited, and reproduced.

New H3 media workflows

Added multiple timed frame injection for First / Last generations.
Added audio-driven video from an uploaded soundtrack or a control video's audio.
Added video-to-audio generation that preserves the source pictures while creating a new synchronized soundtrack.
Added H3 video-to-video editing for the whole frame, inside a mask, or outside a mask, with adjustable denoise and mask strength.
Fixed Omni music and performance references restarting from the beginning in every sequence clip; each window now receives the correct timeline segment while voice references remain reusable.

Memory, performance, and RTX 50 support

Added VRAM-, model-, and resolution-aware H3 window recommendations with a native 14.4-second ceiling and saveable user overrides for proven hardware combinations.
Improved transformer residency, activation workspace, streaming VAE decoding, RAM budgeting, and LoRA fallback behavior for Full and Pruned H3 models.
Added a dedicated RTX 50 / Blackwell runtime with Python 3.11, PyTorch 2.10, CUDA 13, compatible acceleration kernels, automatic migration, startup diagnostics, and one-click repair.
Preserved the established runtime for RTX 20/30/40 systems while applying hardware-specific setup only where required.

Workflow and interface reliability

Added early validation for incompatible H3 media, prompt counts, durations, and sequence settings before expensive model loading begins.
Fixed Omni reference uploads on iPhone and iPad so supported audio files are selectable even when iOS reports unusual file types.
Generation cards now show active generation time in minutes and seconds, excluding time spent waiting in the queue or loading models.
Expanded regression coverage for H3 continuation, reference packing, audio timing, frame injection, video editing, memory recommendations, RTX 50 setup, and the shared multi-window interface.