Repository navigation
Releases
1.3 — One Queue Owns the GPU, Sound Effects, Image → 3D, Film Recipe
New: one job queue owns the GPU
Every tab used to have its own Load button, and starting anything on a second tab meant stopping the first. That scheduling is now the software's job. Every tab submits to one queue, and a runner loads and frees models as it goes. Line up a video, four images, a song and a 3D mesh in any order, then walk away.
- Nothing to load first. Submit from cold; the queue works out which model each job needs.
- A real ETA, per job and for the whole queue. Measured from previous runs of the same kind of work, including model swaps. A job that has never run says so rather than guessing.
- Same-model jobs are pulled forward to save 20 s to 3 min loads, with a cap so nothing waits forever. Reorder by hand with ▲.
- Survives a restart. The queue is saved on every change; interrupted work is re-queued.
- Two lanes. CPU-only work runs beside GPU jobs instead of behind them.
- Priority for interactive work. A chat turn, a transcription or one spoken line jumps the line.
- Pinning. Keep one model resident while you iterate on a tab.
- Cancel means stop. Cancelling a running image, music, sound-effect, 3D or video job also stops the work it had already handed to ComfyUI, so the card is free at once. Only the studio's own work is touched.
New tabs and features
- 🔊 Sound Effects: foley, impacts and ambience from a description (Stable Audio 3.0 Small SFX). Several takes in about a second each.
- 🧊 Image → 3D: one reference image becomes a
.glbmesh (Hunyuan3D 2.1). - 🧹 Reference-clip cleanup: one-click denoise and de-reverb (DeepFilterNet) on the clip you clone a voice from. Runs on the CPU, so it never unloads the voice model.
- 🧠 Language tab writes real files: pages, scripts and documents land in a sandboxed
workspace/folder with a live preview panel. - 🧠 Language tab keeps conversations: every thread is saved on the server, named, listed and resumable after a reload, a second tab or a restart.
- Local LLM context raised from 4k to 32k. Ollama's default had been silently truncating long prompts.
New: Film Recipe
docs/Film_Recipe.docx is a start-to-finish guide to making a narrated short film with the studio, written for any subject: brief, script, timed plan, voice, look bible, keyframes, clips, review batches, sound, assembly and delivery, plus a troubleshooting table.
The README's Demo now leads with Everything Before, a 12-minute narrated sci-fi film planned and directed by Claude Opus 5.5 in Claude Code and rendered on one RTX 4080 SUPER with this studio.
Licence
Local AI Studio is now released under Apache-2.0 (LICENSE + NOTICE). That covers the studio's own code and docs. The models the runbook downloads keep their own licences; see Credits & licensing and the MiniMax H3 notice in the README.
Docs
- Setup runbook: 29 → 32 embedded files, with new phases 5i (Sound Effects, ~3.3 GB) and 5j (Image → 3D, ~6.9 GB).
- All three companion Word docs regenerated for the queue release.
- New screenshots: the job queue, Home, Language, Sound Effects and Image → 3D.
Attached: Local-AI-Studio-Setup-1.3.zip: the setup runbook, the four Word docs (now including the Film Recipe), LICENSE and NOTICE, standalone for anyone who just wants those without cloning the repo.
1.2 — Local Video Generation (MiniMax H3)
New: Video tab — MiniMax H3
Local text-to-video with native stereo audio, generated together in one sampler pass — nothing to mux afterwards. Three input modes:
- Text → video — describe a shot, get picture and sound from the same prompt
- First / last frame — drop one keyframe for image-to-video, or both and H3 interpolates between them
- References — up to 9 images, 3 video clips (each with its own optional soundtrack), and 3 audio clips to carry identity, motion, camera style, or a cloned voice into the result
Runs at H3's native 768p on a single 16GB GPU, streaming the official 21GB int8 weights through VRAM via ComfyUI's per-module offload.
Real numbers, not assumptions
- Recommended length: 5–15 seconds.
- Measured generation time from real runs on this rig: ~3 minutes of generation per second of output video, end-to-end including model load (a 5s clip ≈ 15 min, a 15s clip ≈ 45 min).
- Settings notes from direct benchmarking: run at native 768p or the model stops animating; the official int8 weights beat a smaller GGUF quant on both speed and quality; EasyCache trades audio quality for a speed gain; SageAttention gives no measurable gain on a card this size.
⚠️ Licensing
MiniMax H3's licence excludes the UK, EU, US, and South Korea, and the restriction covers generated output, not only the model weights. Flagged prominently everywhere the Video tab is documented. Apply for a licence if you are in these places: https://huggingface.co/MiniMaxAI/MiniMax-H3/discussions/12
Fixed
- Video tab screenshot previously showed a real list of generated filenames from the developer's own machine — redone to show the panel's genuine empty state.
Docs
Documented across the README, the self-contained setup runbook (h3gen.py now embedded), and all three companion Word docs.
Attached: Local-AI-Studio-Setup-1.2.zip — the setup runbook plus the three Word docs, standalone for anyone who just wants those without cloning the repo.
1.1 — Composer, Sprite Studio, Lullaby, and a navigation fix
New
- Composer — a text brief becomes a fully arranged, mixed, multitrack instrumental. A local LLM plans the musical direction (instruments, key, tempo, structure, chords, mix, automation) from a fixed studio menu; a deterministic engine (
composerkit.py) writes every note and runs the whole mix on General MIDI + FluidSynth, so it never depends on unreliable note-level model output. 3-step wizard: Set up → Arrange (DAW-style clip grid, mixer, piano-roll editing, re-render with no new LLM call) → Export (master, multitrack MIDI, per-instrument stems). - Sprite Studio — one reference image becomes a full style-matched 2D animation set (idle/walk/run/jump/fall/crouch/attack/hurt/death), transparent backgrounds, sprite sheets with engine-ready JSON metadata, per-frame re-roll.
- Lullaby workbench — split any song into 6 stems, then choose an engine: Remix (closest to the original), Piano (rebuilt piano + music-box arrangement), or Melody Match (traces the real sung melody's continuous pitch onto an instrument). Per-track routing lets some stems go through Melody Match while others get a full Piano-style rebuild, mixed together, with an optional ACE-Step polish pass.
- Track Splitter — save any song's 6 separated stems on their own.
- Play all selected — synced playback across selected tracks in the Lullaby/Track Splitter multi-track panels.
Fixed
- A duplicate-declaration bug that broke navigation across the entire app (Sprite Studio and Track Splitter shared JavaScript variable names).
Docs
- Composer, Sprite Studio, Lullaby, and Track Splitter are now fully documented across the README, the setup runbook, and all three Word docs.
- Every screenshot replaced with a real, completed workflow instead of an empty landing state.
Attached: Local-AI-Studio-Setup-1.1.zip — the setup runbook (SETUP_FOR_CLAUDE.md) plus the three companion Word docs, standalone for anyone who just wants those without cloning the repo.
1.0: Add README screenshots and update Voice Studio image
Adds a Screenshots section to the README covering all ten tabs, and swaps the Voice Studio capture to the "Create a voice" wizard state instead of the "Use a voice" loading screen.