3.1.0: one-shot installs from prebuilt engines

Prebuilt engines for a one-shot install on either machine. Neither machine needs the other.

DGX Spark: the carrier ghcr.io/drowzeys/keys-tensorfold-studio:3.1.0 (digest 6f22ddd5) holds the prebuilt CUDA engines for GB10: H3 (1f7ecc75), MiniMax Music 3 (6b714531), Fish Audio Drama 3 (81173e6c) and NVIDIA Nemotron 3.5 Bolt (Mia's recipe 6f61d772). bash oneshot-setup.sh --all installs every feature.

Mac (Apple silicon): the assets below, which the installers download and verify by SHA-256. No Docker needed.

• zig-engine-b19753ab-macos-arm64.tar.gz: TensorFold's native Zig + Metal FastH3 engine, with native LoRA. sha256 775e9163…
• bolt-metal-b4edaea5-macos-arm64.tar.gz: TensorFold's Metal engine for NVIDIA Nemotron 3.5 Bolt (tensorfold-native + libbolt for Apple GPUs; Apache-2.0, see its LICENSE). sha256 58203b65…

bash oneshot-setup.sh --all --app installs every feature your Mac's memory runs. The minimum hardware requirements are at the top of the README.

No model weights are in the carrier or these assets.

2.9.3: FLUX3-K's new file names

• FLUX3-K installs again on a DGX Spark. On 2026-10-17 Black Forest Labs renamed the Klein checkpoints on Hugging
  Face (flux3-image-klein-bf16.safetensors, flux3-image-klein-base-bf16.safetensors) and removed the old
  underscore names, so oneshot-setup-spark.sh --flux3 could not download them. The setup now fetches the new names
  and renames an older install's files in place instead of fetching 22 GB again: they are the same files (the same
  SHA-256 on Hugging Face). The app reads either name, so a Spark updated from the Sparks tab keeps FLUX3-K before
  the setup runs. Nothing changes on a Mac.

• Carrier image unchanged: ghcr.io/drowzeys/keys-tensorfold-studio:2.9.1.

Upgrade

On a DGX Spark with FLUX3-K, in the clone:

git pull
bash oneshot-setup.sh --upgrade --flux3

That renames the two checkpoints to BFL's new names (no download). Without it the app keeps working with the old
names. A new FLUX3-K install (--flux3) now downloads again. On a Mac there is nothing to do.

2.9.2: Music on TensorFold's Metal engine, tutorial for every feature

• Mac: Music runs on TensorFold's native Metal engine. On an M5 or later the Music tab now runs MiniMax Music 3
  on music3-generate built for Metal (Zig and Metal, no Python or MLX; TensorFold pull request #987, commit
  8c9559da), when oneshot-setup.sh --music has built it (scripts/build-music3-metal.sh). On the Mac Studio M5
  Ultra, against the MLX port int8, 60 s hip-hop, seed 7, three runs each: the autoregressive model 69 frames/s
  against 65, the whole song 59.5 to 63.0 s against 62.3 to 63.9 s, peak 17.8 GB against 28.1 GB. Through the app a
  song the model ended at 33.4 s took 35.2 s, load included, at 16.1 GB. Limits: it does not stream yet, so Play
  while it's made is MLX-only (MUSIC3_BACKEND=mlx); an M1 to M4 keeps the MLX port; measured on an M5 Ultra only.
  See Music.

• Tutorial and README brought up to date. The Tutorial tab has new lessons for Voice (Fish Audio Drama 3), Music
  (MiniMax-Music3, including the Metal route), the Sparks tab, Sharp Spec, memory on a DGX Spark and the Library's
  actions, and stale lessons are corrected. TUTORIAL.md covers every feature through 2.9.2. The README's version
  notes now run newest first.

• Carrier image unchanged: ghcr.io/drowzeys/keys-tensorfold-studio:2.9.1 (nothing in it changed; the Metal
  Music engine is built from source on the Mac).

Upgrade

On a Mac, in the clone:

git pull
bash oneshot-setup.sh --upgrade --music

On an M5 or later --music builds the Metal engine (Xcode's Metal toolchain needed; a few minutes) beside the MLX
port; on an M1 to M4 it keeps the MLX port. On a DGX Spark, git pull and restart the app: nothing to install.

2.9.1

Three changes to Music (MiniMax-Music3), one for each machine and one for the Mac's player. Every figure is one run
unless it says otherwise.

• Mac: int8 weights by default. The Studio starts the MLX engine (music3-mlx, now pinned at 6844b413) with
  --quant int8: the autoregressive model, its code-0 head and the depth decoder at 8 bits. On the Mac Studio M5
  Ultra the autoregressive model went from 40 to 65 frames a second and the peak from 27.6 to 20.5 GB; a 60 s song
  took about 63 s through the app (75.8 s in 2.9.0). Against ComfyUI's reference the code-0 logits keep cosine
  0.99994 (bf16 0.99997), 98.0 % greedy agreement, and Whisper's lyric scores over six same-seed songs stayed within
  their noise. MUSIC3_QUANT=none runs the bf16 weights. Limits: measured on an M5 Ultra only; other
  M-series chips are not measured.
• Mac: play while it's made. With the box ticked (the default where the engine streams) the song plays in the tab
  while it is made, from GET /api/music/takes/{id}/live, one stereo 44.1 kHz WAV. The sound is held back until
  playing will not run dry (the engine's measured no-stall start plus 3 s). Through the app, 60 s songs: first sound
  at about 7 s (7.6 and 6.4 s), playing from about 13 s (13.4 s both), no stall (the least buffer 4.0 s), made in
  63.4 and 61.0 s; a 30 s song played from 8.1 s. The kept take is the engine's finished WAV, the same samples that
  were played. Limits: a Mac only (no route on a Spark streams); only 30 s and 60 s were measured through the app, and
  the wait grows with the length (about 56 s for a 300 s song, from the engine's own figure); nobody has listened for
  seams. Details in Music.
• Music on a DGX Spark: TensorFold's CUDA engine for MiniMax Music 3 (MiniMax-Music3). The Music tab on a Spark now
  runs music3-generate, TensorFold's native CUDA port (branch music3-cuda, 6b714531: Zig + CUDA, no Python at run
  time), wherever it is built and the weights are there; otherwise it falls back to ComfyUI's nodes as in 2.9.0.
  MUSIC3_BACKEND=cuda|comfy chooses; the tab's state says which route runs and, after a fall-back, why.

  • int8 weights by default, made on the GPU as the engine loads (no cache file): the autoregressive model, its
    head and the depth decoder at 8 bits. MUSIC3_WEIGHTS=bf16 runs the checkpoint's own weights. It reads
    Comfy-Org's three files as they are (~/spark-studio-models/MiniMax-Music3-Comfy, MUSIC3_MODEL names another).

  • Same lifecycle as the Mac engine: a persistent server, loaded on the first song, holding
    ~/studio-gpu.lock.d while loaded, unloaded after five idle minutes and before any other job, under the
    memory kill-guard; a studio ComfyUI that holds models is stopped before it loads. Never beside H3, Drama 3 or
    ComfyUI's models.

  • Installed by bash oneshot-setup-spark.sh --music (and --upgrade where the weights are): spark/build-music3.sh
    takes the prebuilt binary from the carrier image, or builds it with nvcc 12.9+ and Zig 0.17 (50 s on a GB10).

  • Measured on one GB10 (int8, hip-hop lyrics, seed 7, 30 steps; one run each; memory is MemTotal - MemAvailable
    above the idle machine):

    Song	TensorFold CUDA (int8)	ComfyUI's nodes (2.9.0)
    30 s, engine	60.3 s	131.0 s
    60 s, engine	123.9 s (AR 73.9 s at 20.4 frames/s, DiT 46.4 s, DAV 3.6 s)	260.6 s
    120 s, engine	252.0 s	527.6 s
    60 s through the app	149.2 s for the song (load 37.3 s, AR 68.2 s, DiT 40.4 s, DAV 3.3 s; the model ended it at 55.2 s), 171.4 s with Whisper's check (97 %, 164 of 169 words)	270.6 s for the song (ComfyUI started for it; AR 166.5 s, DiT 94.3 s, DAV 4.0 s; 59.99 s made), 297.0 s with Whisper (96 %, 162 of 169), 29.4 GB at the peak: the same run of the 2.9.1 app with MUSIC3_BACKEND=comfy
    Peak memory, 60 s	17.9 GB (18.7 GB through the app)	29.3 GB

    bf16 weights: a 60 s song in 194.3 s and 25.5 GB. Parity against ComfyUI's nodes and Whisper's lyric count
    passed on the engine branch (docs/recipes/minimax-music3-cuda.md there).

  • Honest limits. No streaming on the Spark: a song is heard when it is done. Built and measured on a GB10
    only (the build refuses another GPU unless MUSIC3_ANY_GPU=1; elsewhere use MUSIC3_BACKEND=comfy). Loading
    takes about 37 s each time the engine starts (int8 is made at load), so the first song after five idle minutes
    pays it again. Songs over 120 s are not measured on this engine; the memory estimate goes on at the measured slope.

• Carrier image ghcr.io/drowzeys/keys-tensorfold-studio:2.9.1 adds the prebuilt Spark music3-generate
  (spark-music3/) beside the H3 engine.

Upgrade

On a Mac or a DGX Spark, in the clone:

git pull
bash oneshot-setup.sh --upgrade --music

--music installs the Music engine for that machine (the MLX port at 6844b413 on a Mac; on a Spark music3-generate from the carrier image ghcr.io/drowzeys/keys-tensorfold-studio:2.9.1) and the 22 GB of weights where they are missing. Leave --music out to upgrade without Music. Then restart the app.

2.8.0: FLUX3-K picture models and edits

2.8.0: pick the picture model: Qwen 2.1 Turbo, FLUX3-K Fast or FLUX3-K Full

• FLUX3-K (Black Forest Labs' FLUX 3 Image Klein) joins Qwen-Image-2.1-Turbo wherever the Studio makes a picture: Image only, scouts, last frames, and Edit a picture, with up to nine more reference pictures. Fast is the distilled checkpoint (16 steps); Full is the base one (50 steps, CFG 3.5). BFL's prompt upsampler writes the caption first.
• DGX Spark: runs on ComfyUI's own FLUX 3 nodes, and ComfyUI moves to 5a665890.
• Mac: runs on flux3-mlx in its own environment.
• Install: oneshot-setup.sh --flux3 (41 GB; accept the FLUX Kommunity License on Hugging Face first).

Measured at 1344x768, bf16, with the model loaded. Picture times:

	DGX Spark	M5 Ultra
Qwen Turbo		9.7 s
FLUX3-K Fast	10.5 s	14.9 s
FLUX3-K Full	57.5 s	53.6 s

Edit times, Qwen / Fast / Full:

	DGX Spark	M5 Ultra
Edit	24 / 37 / 139 s	10 / 37 / 240 s
Two pictures	23 / 44 / 188 s	13 / 42 / 427 s

On a Mac, use Fast for edits.

Carrier image: ghcr.io/drowzeys/keys-tensorfold-studio:2.8.0, digest sha256:44dba56cb8e316cdc35a3e227f1d5109901ad9209a0cdb408e3ef27e5a2265be, for linux/arm64 and linux/amd64. It now holds the DGX Spark's LoRA engine (h3-cuda-lora 1f7ecc75), so a Spark install no longer builds it from source.

Upgrading: run git pull, then bash oneshot-setup.sh --upgrade; add --flux3 to fetch the FLUX3-K weights.

Licences: FLUX 3 Image Klein is under the FLUX Kommunity License: non-commercial use, and commercial use of the outputs only for Qualifying Users (under US$5M revenue). Qwen-Image-2.1 keeps the Qwen Research License.

What's Changed

• 2.7.3: built with integrated Fish Audio Drama 3 (name, title, credits) by @drowzeys in #34
• 2.7.4: Voice, with Fish Audio Drama 3 on the Mac by @drowzeys in #35
• 2.7.5: Voice on the DGX Spark (TensorFold native CUDA Drama 3) by @drowzeys in #36
• 2.7.6: Mac Voice plays while made and checks its own words; public engine links; --voice installer step by @drowzeys in #37
• 2.7.7: Mac Drama 3 as a persistent server (first sound ~0.25 s); dialogue a turn a line by @drowzeys in #38
• 2.7.8: CUDA takes with missing words are redrawn; the Mac engine hears a clip to clone by @drowzeys in #39
• 2.8.0: pick the picture model: Qwen 2.1 Turbo, FLUX3-K Fast or FLUX3-K Full by @drowzeys in #40

Full Changelog: v2.7.2...v2.8.0

2.7.2: typing a custom size

With Sam Wasserman (wassermanproductions.com): Slate, the Wasserman Sharp Spec, its QA gate, the 80s Horror LoRA and Motion Previs Studio are his work. And with the LoRAs' authors: Jojocodex, neph1, shamanic, pablodawson, Ashmotv, matlod, alvdansen, Akatz Labs (akatz-ai), vpakarinen, JOKER141, adf99, DiffSynth-Studio, SOLRICKS, dntpi, KennethFal, akhaliq.

Engine credits as before: TensorFold (Ash Hart and contributors), MiniMax H3 (MiniMax), FastH3 (FastVideo), Qwen-Image-2.1 (Qwen), minimax-h3-mlx (mrbizarro), h3.c (Salvatore Sanfilippo), mflux (Filip Strand), LightX2V, linjian257. See CREDITS.md.

New in 2.7.2: typing a custom size

Fixed: the Width and Height boxes of a custom resolution, and the box for a custom number of passes, took one digit at a time. Every key redrew the form and the box under the cursor with it. The value still changes with every key; the form is redrawn once the box is finished (Enter, Tab or a click elsewhere).

Checked: the seed. The same prompt and settings at seeds 11 and 12 give two different clips on every route:

Route	Frame 10, mean difference
Mac, Zig engine	64 / 255
DGX Spark, TensorFold int8 engine	29 / 255
DGX Spark, int8 engine with a LoRA merged in	35 / 255

The Create tab, each Production segment and each Slate shot keep a seed of their own, and the Sharp Spec sets every segment or shot to 42. Change the seed where the clip is rendered from.

Upgrade

Page and test changes only; the engine is unchanged. Update the app on the head DGX Spark and the Mac, then press Update and start for each worker on the Sparks tab.

Wasserman Slate AI Studio 2.7.1

With Sam Wasserman (wassermanproductions.com): Slate, the Wasserman Sharp Spec, its QA gate, the 80s Horror LoRA and Motion Previs Studio are his work. And with the LoRAs' authors: Jojocodex, neph1, shamanic, pablodawson, Ashmotv, matlod, alvdansen, Akatz Labs (akatz-ai), vpakarinen, JOKER141, adf99, DiffSynth-Studio, SOLRICKS, dntpi, KennethFal, akhaliq.

Engine credits as before: TensorFold (Ash Hart and contributors), MiniMax H3 (MiniMax), FastH3 (FastVideo), Qwen-Image-2.1 (Qwen), minimax-h3-mlx (mrbizarro), h3.c (Salvatore Sanfilippo), mflux (Filip Strand), LightX2V, linjian257. See CREDITS.md.

New in 2.7.1: a LoRA on a DGX Spark renders on TensorFold's int8 engine

The TensorFold node (TensorFoldH3Blocks) takes a LoRA's patches on the block projections off ComfyUI. It merges them into FastH3's bfloat16 weights the way ComfyUI does, and the engine quantizes them to int8 through tf_h3_set_weight (engine drowzeys/TensorFold branch h3-cuda-lora @ 1f7ecc75). The LoRA's other keys stay ComfyUI's. The saved .tf-int8 copy is never changed.

What	2.7.0 (ComfyUI's bf16 blocks)	2.7.1 (int8 engine)
Largest LoRA clip taken	1280x736, 124 frames	1376x768, 243 frames (10 s Sharp Spec)
864x480, 124 frames	140 s, 75.9 GB	114 s (+12.5 s merge the first time), 57.6 GB
1376x768, 124 / 192 / 243 frames	refused	62.3 / 66.5 / 69.7 GB

The memory figures are the lowest point of available memory on a 124.5 GB worker Spark, measured from below under a second watchdog at 25 GB. They match the engine's own figures without a LoRA.

Same seed, The Wasserman 80s Horror at 1.0:

• The engine's clip and ComfyUI's bfloat16 clip show the same shot: pixel correlation 0.80, sound 0.89.
• That is closer than the two engines are to each other without a LoRA (0.55).
• Both are far from the clip without the LoRA (0.43 to 0.49).

Unchanged: Ref2VA, MiniMax H3's own transformer (--base h3) and Turbo keep ComfyUI's blocks. A Spark whose engine predates 2.7.1 renders a LoRA on the bfloat16 blocks with the old limit, and says so.

Upgrade

1. Run bash oneshot-setup-spark.sh --upgrade on the head DGX Spark. It rebuilds the engine at the new commit (spark/build-engine.sh, about a minute).
2. Copy libtf_h3.so to each worker, or run spark/build-engine.sh there.
3. Press Update and start for each worker on the Sparks tab.

Nothing changes on a Mac.

Licences are unchanged: Qwen-Image-2.1 is under the Qwen Research License (non-commercial); MiniMax H3 and LoRAs made from it are under the MiniMax H3 Community License Agreement, with its territory limits.

Wasserman Slate AI Studio 2.7.0

With Sam Wasserman (wassermanproductions.com): Slate, the Wasserman Sharp Spec, its QA gate, the 80s Horror LoRA and Motion Previs Studio are his work. And with the LoRAs' authors: Jojocodex, neph1, shamanic, pablodawson, Ashmotv, matlod, alvdansen, Akatz Labs (akatz-ai), vpakarinen, JOKER141, adf99, DiffSynth-Studio, SOLRICKS, dntpi, KennethFal, akhaliq.

Engine credits as before: TensorFold (Ash Hart and contributors), MiniMax H3 (MiniMax), FastH3 (FastVideo), Qwen-Image-2.1 (Qwen), minimax-h3-mlx (mrbizarro), h3.c (Salvatore Sanfilippo), mflux (Filip Strand), LightX2V, linjian257. See CREDITS.md.

New in 2.7.0: the Sparks setting, one to eight DGX Sparks for a job

What	In short
Sparks tab	Add a worker DGX Spark by the address you would give ssh; check it, update its Studio code and start its app, stop it. Each row shows version, memory guard, memory and disk, and what it renders
Sparks chooser, 1 to 8	At the top of the Queue, in Create, on the Production board and on a Slate shot; sparks: N in the API. The app uses as many as are up and able to do the job and says so ("5 of 8 available")
Several clips	One clip per Spark, in parallel, landing in the head's Library and project as if rendered there, with seed, settings and Spark on each take
One clip, long videos	One clip stays on one Spark. A long video stays on the head: each piece opens on the last frame of the one before
LoRA and reference clips	Only to a worker that has them and whose memory guard accepts the clip
How it connects	Each worker's app listens on its own 127.0.0.1; the head opens one ssh tunnel to it. No sudo, no shared folders. A worker silent for 60 s during a job is marked down and its clip rendered elsewhere

Measured

The same eight clips (56 frames, 864x480, FastH3 8 passes, int8 engine) through the app, one run each:

Sparks	Wall time	Against one Spark
1	263.2 s	1.00x
2	129.8 s	2.03x
3	97.5 s	2.70x
4	69.0 s	3.81x
5	65.0 s	4.05x

Five is little better than four for eight clips: they go in two rounds either way. A three-minute music video of 22 ten-second clips on five Sparks, made with a hand-written driver before this setting existed, reached 4.09x.

Not built: one clip split across several Sparks (measured at 72.2 s on one against 40.9 s on four, but the installer does not build that engine and the app does not run its ranks; docs/multi-spark.md lists what is missing). Not tested: six to eight Sparks (five exist here), LoRA and reference clips on real workers, Production and Slate clicked through with real workers. If the head app restarts during a job, a worker's finished clip is not fetched. A Mac shows 1: multi-Spark rendering is for DGX Sparks.

Upgrade

bash oneshot-setup-spark.sh --upgrade on the head DGX Spark, then add each worker on the Sparks tab and press Update and start (a worker needs the Studio and its models installed once with oneshot-setup-spark.sh). bash oneshot-setup.sh --upgrade on a Mac. The carrier image stays ghcr.io/drowzeys/keys-tensorfold-studio:2.6.5.

Licences are unchanged: Qwen-Image-2.1 is under the Qwen Research License (non-commercial); MiniMax H3 and LoRAs made from it are under the MiniMax H3 Community License Agreement, with its territory limits.

Wasserman Slate AI Studio 2.6.6

With Sam Wasserman (wassermanproductions.com): Slate, the Wasserman Sharp Spec, its QA gate, the 80s Horror LoRA and Motion Previs Studio are his work. This release comes out of two music videos made through his Slate. And with the LoRAs' authors: Jojocodex, neph1, shamanic, pablodawson, Ashmotv, matlod, alvdansen, Akatz Labs (akatz-ai), vpakarinen, JOKER141, adf99, DiffSynth-Studio, SOLRICKS, dntpi, KennethFal, akhaliq.

Engine credits as before: TensorFold (Ash Hart and contributors), MiniMax H3 (MiniMax), FastH3 (FastVideo), Qwen-Image-2.1 (Qwen), minimax-h3-mlx (mrbizarro), h3.c (Salvatore Sanfilippo), mflux (Filip Strand), LightX2V, linjian257. See CREDITS.md.

New in 2.6.6

Two three-minute music videos were made through the app on up to five DGX Sparks. This release fixes what they ran into.

What	In short
Memory guard (DGX Spark)	A LoRA clip at 1376x768 took a Spark down. A job is now checked before it is queued and again before anything loads, from measured peaks; a LoRA clip above the largest one measured (1280x736, 124 frames) is refused with the size that fits. A watch beside every render server reads memory twenty times a second and stops the render under a floor of a tenth of the machine
Edit a picture	Any seed gives an edit (the seed that made the source used to return a sharpened copy); up to nine further pictures; the result has the source's own size. A given seed gives a different picture than in 2.6.5
Clip lengths	Any of 21 lengths from 22 to 362 frames, in place of 5, 8, 10 and 15 s
Prompt check	Negations backfire ("no photographs" produced photographs). The page lists negated things before queueing and never changes the prompt; the app's own added lines are now positive
Shot check	After a render: were the first and last keyframe held, is there a hard cut inside the clip, and which is the longest stretch without one; optional retake at the next seed
Slate for music videos	Tempo and bar grid, shot slots in bars, a LoRA per shot (through the guard), a change action for characters and places, the shot's intent kept out of the prompt, takes from another machine imported into a shot
Also	720p on a DGX Spark without ffmpeg; a QA gate that says when it did not run and can gate a worker's clip from a head node; an API that keeps answering while a LoRA job starts

Measured

Test	Result
The job that took a Spark down (LoRA, 1376x768, 5 s)	Refused by the app and by the scripts before anything loads
Run-time floor, raised on purpose on a safe clip	Server stopped in 0.5 s, job failed with the reason
LoRA clip memory, 864x480 to 1280x736, 22 to 124 frames	74.8 to 76.6 GB taken; int8 engine 57.6 to 62.2 GB at 124 frames
Edit at the source's own seed	Correlation with the source 0.90 before, 0.36 after (DGX Spark); 0.97 to 0.15 (Mac)
Hard cuts in the first production's 32 clips	12 of 12 found, none invented in 7081 frame pairs

Not established: why the Spark went down. Its log stops with no out-of-memory kill and no panic. Not tested: the new pages in a browser beyond a load, a Mac under 96 GB, reference clips at the new lengths, LoRA clips above 1280x736 (on purpose), the First AD's new actions with a live brain. A dissolve to a close-up is not detected, and no wording or setting tried stops the model from cutting inside a clip. The memory guard does nothing on a Mac.

Upgrade

bash oneshot-setup.sh --upgrade on a Mac, bash oneshot-setup-spark.sh --upgrade on a DGX Spark. The carrier image stays ghcr.io/drowzeys/keys-tensorfold-studio:2.6.5.

Licences are unchanged: Qwen-Image-2.1 is under the Qwen Research License (non-commercial); MiniMax H3 and LoRAs made from it are under the MiniMax H3 Community License Agreement, with its territory limits.

Wasserman Slate AI Studio 2.6.5

With Sam Wasserman (wassermanproductions.com): Slate, the Wasserman Sharp Spec, its QA gate, the 80s Horror LoRA and Motion Previs Studio are his work. And with SOLRICKS, whose 3D Animation Style LoRA this version puts to use, and the LoRAs' other authors: Jojocodex, neph1, shamanic, pablodawson, Ashmotv, matlod, alvdansen, Akatz Labs (akatz-ai), vpakarinen, JOKER141, adf99, DiffSynth-Studio, dntpi, KennethFal, akhaliq.

Engine credits as before: TensorFold (Ash Hart and contributors), MiniMax H3 (MiniMax), FastH3 (FastVideo), Qwen-Image-2.1 (Qwen), minimax-h3-mlx (mrbizarro), h3.c (Salvatore Sanfilippo), mflux (Filip Strand), LightX2V, linjian257. See CREDITS.md.

New in 2.6.5, on a Mac and on a DGX Spark

What	In short
3D Animation Style (SOLRICKS)	Installs and renders. Its file uses kohya's own layer names, which the hub refused; the hub and the Mac converter now read them. Start the prompt with 3dsrx; 0.8 for people from text, 1.0 for animals and places, 0.4 from a first frame (the author's settings)
What this machine can use	22 of the 22 presets on a DGX Spark, 20 on a Mac (Upscale and Remove a person are DGX Spark only)
Essential LoRA package	Third build: 14 LoRAs per platform (2.1 GB) in one download, now with Multiple Angles. The other 8 come from their authors' pages; 3D Animation Style and Keyframe Animation need their authors' approval of your Hugging Face account
Carrier image	ghcr.io/drowzeys/keys-tensorfold-studio:2.6.5: the engines of 2.6.1 with this release's render scripts

Measured

Test	Mac Studio M5 Ultra	DGX Spark
3D Animation Style at 1.0, same seed with and without, FastH3 8 passes 864x480, 5 s	83 s against 84 s; the clip becomes a 3D animated film (PSNR 16.0 dB between the two)	152 s against 149 s; the same (14.8 dB)
Essential bundle into an empty folder	not run	22 of 22 present in 200 s (14 from the package, 8 from their authors)

Not measured: the Mac archive of the third package build unpacked on a Mac, and 3D Animation Style from a first frame.

Upgrade

bash oneshot-setup.sh --upgrade on a Mac, bash oneshot-setup-spark.sh --upgrade on a DGX Spark, then bash oneshot-essential-bundle.sh for the LoRAs.

Licences are unchanged: Qwen-Image-2.1 is under the Qwen Research License (non-commercial); MiniMax H3 and LoRAs made from it are under the MiniMax H3 Community License Agreement, with its territory limits.