Imagination-Studio-by-keys-Wasserman

Native Windows, the Console look with the simple studio, and Quick Edit. The Mac and Spark engines are the same as 4.2.1's. To update, git pull and restart. On Windows, run the setup again, since 4.4.0 adds its own setup scripts.

• Windows, natively (#200, #206, #207, #211, #212; ivberg's #198 and #210 carried): the whole Studio runs on Windows on Arm (RTX Spark N1X), with no WSL. It includes ComfyUI, TensorFold 1.2.2 engines, Library actions, Character Swap, foley, Finish, LoRA training and Music 3's own engine. Native video fits within Windows' 30 GB. The WSL version is gone.
• The Console look (#219, MiaAI-Lab; #197): Light, Dark, OLED and Match the computer themes, a Text size setting, and the screens rebuilt.
• The simple studio (#220, MiaAI-Lab): a simpler starting view, with the next actions under every result.
• Quick Edit (#208, MiaAI-Lab): change a picture by saying what to change, in one box.
• No sound (#205, MiaAI-Lab): a clip can be made without sound, from step 3.
• Access (#218, MiaAI-Lab): you can choose to trust your own tailnet and home networks. Off by default.
• Library (#222, nickmontag's #221 carried): players load only when they come near the screen, 60 tiles show per page with Show more, and the page redraws only for a new or deleted take.
• Finish (#216): Premiere export works everywhere (the converter is fixed on Windows), and Windows renders frame-exact.
• keys-HDR LORA (#190): an HLG HDR finish for any clip.
• Two GPUs in one machine (#191, davidlowjw's #184 carried) and paired cards in the web UI (#194, davidlowjw's #192 carried).
• Sparks: live readings are drawn in place, so a playing clip no longer restarts every 4 s (#193, davidlowjw, fixes #186). Cancelled clips are no longer sent to workers (#209, newjordan).
• For apps that drive the Studio: SPARK_PROGRESS_FILE writes a clip's progress to a file, for a workload governor (#189, benbuschmann's #182, plus a Cancel fix). Job previews give a clip's live preview over the API, opt-in (#217).
• Home (#196, wassermanproductions): FLUX3 Klein and YuE2 chips in the composer.
• RTX Spark N1X: the GPU carve-out is used as the memory pool, and STUDIO_MEMORY_GUARD=warn flags instead of refusing (#199). Speed and LLM measurements are in the docs (#202, #203, #204).
• Fixes: a docs table row (#188, halr9000), and two flaky tests (#180, #181).

Imagination Studio 4.3.1 runs on one to four local NVIDIA GPUs on one Linux machine, and raises a Mac's default install to 96 GB. The engines are the same as 4.2.1's.

New

• Local NVIDIA GPU workers (#177, akol1's #168 carried onto 4.3.0, thank you @akol1).
  • On a workstation with up to four NVIDIA cards, the Studio runs on one card and starts a worker on each other compute card.
  • Machine → Sparks shows each card's VRAM, utilization and job.
  • Card pinning: workers are pinned by card UUID, so a card whose order changed is caught rather than used by mistake.
  • Shared RAM: jobs wait their turn for the machine's RAM. A job that can never fit is refused at once, and the refusal names what holds the memory.
  • Which card the Studio uses: with two or more compute cards, set STUDIO_GPU to the Studio's own card; the launch script asks for it. A display card is only used when you say so.
  • Limit: a fifth card is refused.
  • See Local GPUs.

Changed

• A Mac needs 96 GB for the default install (video, voice, music, Bolt), up from 64 GB (#178).
  • Pictures only: bash oneshot-setup.sh --image-only still needs 48 GB. A fresh install on a smaller Mac stops before downloading anything and offers it.
  • Upgrading a 64 to 95 GB Mac: --upgrade keeps what's installed, updates it to this version without fetching weights, and says the new minimum. Video keeps running there; the Voice and Music tabs now say they need 96 GB.
  • Unchanged: the 128 GB (MiniMax H3 Turbo, Ref2VA) and 192 GB (Upscale, Remove a person) tiers.

Install or update

• Update: git pull, then run --upgrade as usual and restart the app. Installed engines whose commit didn't change are kept.
• Prebuilt parts: unchanged. The carrier image stays ghcr.io/drowzeys/keys-tensorfold-studio:4.2.1, and the Mac Bolt archive stays the v4.2.1 asset.

Tested

• CI on #177, #178 and this release.
• #177 on a DGX Spark:
  • The full suite: 1,359 passed, 0 failed.
  • Simulated cards:
    • a display card plus a compute card, in either order;
    • two compute cards with and without STUDIO_GPU;
    • four cards, plus a refused fifth;
    • every shutdown left nothing behind.
  • RAM admission against real memory.
  • A real 30 s song rendered from the branch.
  • Not tested by us: two physical cards rendering at once. akol1 reports a two-card run.
• #178 on Linux and a Mac: new installer tests fake a Mac at 32 to 128 GB, plus the Voice and Music refusals at 64 and 95 GB.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PfjmufXm4L7qsVNFyRw9ZK

Imagination Studio 4.3.0 adds Apps, every tool of the Studio as a one-click app. Songs can now be cleared to the trash, and the Library's automatic captions no longer load a Bolt that can't see. The engines are the same as 4.2.1's.

New

• Apps (#176, Sam Wasserman): a new room at the top of the sidebar.
  • Nine kinds (Cinema, Characters, Camera, Effects, Styles, Sound, Finish, Filmmaker Suite and Studio), with a search box. A DGX Spark lists 61 apps.
  • Pre-filled forms: each app opens the Studio's own form with its model, LoRA or start already chosen. It adds no new engine or server route.
  • Nothing to set up first:
    • A LoRA that isn't installed yet opens Looks on that LoRA.
    • A Library action (Character Swap, Restyle, Fix, Extend, Remove a person, Upscale, Captions, Turnaround) first asks for a clip or picture from your newest takes.
    • An app this machine can't run says why on its card.
  • Character Swap · Openjutsu: the swap is named for what it is, Comfy Org's Openjutsu made inside the Studio, fully local. Search for Openjutsu or Genjutsu to find it.
  • Shortcuts and phones: every app is in ⌘K. On a phone, Apps sits in the More sheet.
  • Marks and art: YuE2's mark appears beside its name in Music and in Apps; it's recorded in THIRD-PARTY-MARKS.md. There's new key art for Captions, Titles, Restyle, Fix and Extend.

Fixed

• Songs can be cleared to the trash (#170, Sam Wasserman).
  • What goes with it: a cleared song takes its record, and for a YuE2 song its score folder, to the trash. Put back restores every file.
  • What stays: a song stays where it is while a film's soundtrack or a queued export or finish names it, or while it's still being made. The skipped list says why.
  • Before: a finished song could never be cleared.
  • Test fix: a test that counted other jobs' engine unloads no longer fails in some full runs.
• Library captions don't load a Bolt with no vision (#175). A Mac's Metal Bolt reads text only. The automatic clip captions used to load it for every new clip, get no answer, and keep it from unloading. Now the captioner skips a Bolt without a vision tower, and the Studio's Bolt unloads on its usual idle timer.
  • If you turned off Settings → "Describe new clips automatically" on a Mac because of this, you can turn it back on.

Install or update

• Update: git pull, then reload the page; the 4.2.1 engines are unchanged.
  • On a DGX Spark or a Linux workstation, --upgrade is still safe to run: bash oneshot-setup-spark.sh --upgrade or oneshot-setup-rtx-6000.sh --upgrade. It keeps every installed engine whose commit didn't change.
  • Restart the app after either.
• New install: clone the repo and run bash oneshot-setup.sh.
• Prebuilt parts: unchanged. The carrier image stays ghcr.io/drowzeys/keys-tensorfold-studio:4.2.1, and the Mac Bolt archive stays the v4.2.1 asset.

Tested

• CI on each pull request and on main.
• New tests/test_apps_page.py: rooms, kinds, LoRA use, install and unsupported cases, the take picker for Library actions, search and ⌘K.
• New tests/test_clear.py cases for songs.
• On a DGX Spark: 61 of 61 apps ready; LoRA apps open Create with the look on; Upscale lists recent clips; the phone layout at 375 px has no sideways scroll.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PfjmufXm4L7qsVNFyRw9ZK

Imagination Studio 4.2.1 removes build-machine paths from the prebuilt engines. The engines' code and output are unchanged.

What changed

• Engines: the H3 engine (libtf_h3.so) and Drama 3's drama3-generate and libtensorfold_drama3.so for the DGX Spark carried build folders in their debug sections, unchanged since 3.3.0. So did the Mac's Bolt tensorfold-native in 4.2.0. All four are now stripped of debug information. Their code, data and CUDA kernels are byte for byte the same; only a few header bytes differ.
• Carrier: ghcr.io/drowzeys/keys-tensorfold-studio:4.2.1 holds the stripped H3 and Drama 3. The Mac Bolt archive on this release is the stripped one.
• Guard: scripts/build-carrier.sh now strips every engine it packs and refuses a payload that names the building machine's home folder or user. A new test checks this.

Same output

Old and new builds were run alternately (old, new, new, old) on the same machine and seeds:

Update

git pull, then --upgrade as usual, and restart the app. An installed engine is kept while its commit is unchanged. To take the stripped builds, remove only each engine's version marker before --upgrade: on a DGX Spark rm $PREFIX/tf-h3/commit $PREFIX/drama3/commit $PREFIX/drama3/BUILD, then run bash scripts/setup-voice.sh (or oneshot-setup-spark.sh --voice) for Drama 3. On a Mac, rm $PREFIX/bolt-metal/COMMIT. Don't delete the drama3 folder: it also holds the converted codec link, codec-v1.safetensors. If it was deleted, setup-voice.sh converts the codec again in a few seconds.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PfjmufXm4L7qsVNFyRw9ZK