Repository navigation
Conversation
593b091 to
4832943
Compare
A web app (FastAPI and plain JS pages) that queues picture and video jobs on a Mac (Metal) or a DGX Spark (CUDA): Qwen 2.1 Turbo pictures and edits, MiniMax H3 and FastH3 clips with sound, reference pictures, videos and audio, long chains, the Production board and Editor, the Wasserman Slate workflow (a port of Sam Wasserman's Slate), the LoRA hub, Library actions, the QA gate and shot check, the memory guard and multi-Spark rendering. The render scripts under scripts/ and spark/ drive TensorFold's H3 and Qwen-Image families; tests run without a GPU.
Install and use, measured receipts, the agent rules, the credits of every author the Studio builds on, and the licence notices; no model weights ship with the Studio.
…Previs Studio The tab takes a shot apart (camera move in words, pose and OpenPose skeleton, depth, edges, line art, motion mask) in its own environment, and hands one layer to the create wizard as <Video 1> of a MiniMax H3 reference clip, with the control-video prompt, the pace rules (Impacts, Flow, Automatic) and the measured sound-and-movement flag. Motion Previs Studio is Copyright 2026 Sam Wasserman, Apache-2.0; its NOTICE, the third-party notices and the file headers are kept.
…urbo FLUX3-K Fast (distilled, 16 steps) and Full (base, 50 steps) as picture models for images, scouts, last frames and picture edits with up to nine further references. On a DGX Spark they run inside ComfyUI's FLUX 3 workflow under the memory guard, on a Mac on flux3-mlx in an environment of their own; BFL's prompt upsampler writes the caption first. FLUX Kommunity License; no weights ship.
4832943 to
e2b4293
Compare
|
Thanks @drowzeys for the image-model chooser and the Fast render measurements of 10.3 to 10.7 seconds and 14.9 seconds. Your separation of installed models and refusal of unsupported edit options is useful for the app's design. As in #941, we're building the media workspace in TensorFold's app itself, with credit for these ideas as its stages land. We'll close #943 with its corresponding stage. The supplied FLUX renders use other runtimes, so they do not establish a native TensorFold FLUX family. This follows the app's release schedule; no FLUX feature is promised for the 1.1.0 engine release. No further changes are requested on this PR. Posted by Ash's AI Assistant Synthena |
|
Thank you. Our plan for the whole Studio stack is in #941: we're folding it into the TensorFold app with your authorship kept. We'll keep this open and link the pieces here as they land. - Synthema, Ash's AI Assistant. |
Stacked on #942 (
upstream/studio-previs), which is stacked on #941; merge after both.Commits of this PR only: drowzeys/TensorFold@upstream/studio-previs...upstream/studio-flux3k (one commit).
The Studio now makes its pictures with Black Forest Labs' FLUX 3 Image Klein as well as Qwen-Image-2.1-Turbo, on both machines. A person picks the picture model for an image, the scouts a clip starts from, a last frame, and a picture edit with up to nine further references. Qwen 2.1 Turbo stays the default.
Measured at 1344x768 with the model loaded, bf16, one run each (
studio/docs/flux3-k.md):Why it matters for TensorFold: a second image family in the same app, chosen per job, with the result feeding TensorFold's video engine as a first frame, a last frame or a reference.
What it adds
studio/app/picture_models.py: FLUX3-K Fast and Full join Qwen 2.1 Turbo inIMAGE_MODELS, offered only where installed (their ComfyUI files on a Spark, flux3-mlx and BFL's checkpoints on a Mac).studio/app/picture_edit.pyandstatic/edit.js: the model choice in Edit a picture; image LoRAs, the camera ring and Qwen's 40-step base model stay Qwen's, and a FLUX3-K edit refuses them.create.jsandtutorial.js(the chooser itself is in the base PR and shows once a second model is installed).flux3_generate.pyandflux3_edit.pyon flux3-mlx;scripts/setup-flux3.sh(oneshot-setup.sh --flux3);IMAGE_MODEL=flux3k-fast|flux3k-fullinscripts/image.shandscripts/edit.sh.spark/flux3_workflows.py, the FLUX3-K ComfyUI workflows (text to image withTextGeneraterunning BFL's upsampler with their settings; the edit withReferenceMultiLatentand the upsampler's i2i prompt), dispatched fromspark_generate.pyandspark_edit.py;oneshot-setup-spark.sh --flux3(41 GB).tests/test_image_models.py, six tests: only installed models are offered and asked for; an edit goes to the picked model (and a FLUX3-K edit refuses Qwen's LoRAs, angles and base model); the ComfyUI workflows FLUX3-K becomes on a Spark, text to image and edit, follow BFL's layout; the upsampler is shown the pictures on one canvas; an unknown model is refused by the script.docs/flux3-k.md, the README's Image only and Edit a picture parts,NOTICE,CREDITS.md,AGENTS.md.How it relates to the engine and to
app/FLUX3-K does not run on TensorFold yet: on a Spark it is ComfyUI's FLUX 3 implementation, on a Mac felix-ab's flux3-mlx. It sits behind the same
IMAGE_MODELswitch as Qwen, so a TensorFold FLUX 3 family can take its place with no change to the app or the page. In the desktop app it is part of the embedded Studio (see #941).Tests
The six new tests pass on both; the failures are the same environment ones as in #941.
Lean check
python3 tools/zig/lean_check.py: 1941 files, 517 problems, unchanged (zig/andtools/zig/only). Overstudio/with the same rules: 21 findings, one more than #942:"text_encoder/*"in the Python insidesetup-flux3.sh, a glob the checker reads as a block comment.Not tested
What comes next
Licences
🤖 Generated with Claude Code
https://claude.ai/code/session_01PfjmufXm4L7qsVNFyRw9ZK