osiworx-pack

osiworx-pack

NO QUALITY LOSS: MiniMax H3 output is BIT-IDENTICAL with and without chunking

The LayerStream: H3 Chunking node does not approximate anything. It runs the same math in slices. We compared the decoded video frame by frame against stock ComfyUI: every frame identical (PSNR = infinity), verified for 124 frames at 6 steps and for the full 15 s / 1 MP clip (362 frames, 2 and 6 steps). The audio track was compared byte for byte on a short clip (56 frames, 6 steps): identical (a 15 s audio comparison is still to be run). If you see a difference in an H3 result, that is a bug: please report it.

Krea 2: with a single chunk the result is also bit-identical (it is the same code path). With several chunks on a bf16 model the matrix multiplications round differently at different slice sizes, so pixels are not bit-identical (38-41 dB against stock; the images look the same). We have not yet tested whether int8/fp8 Krea2 files behave like H3 (exact) or like bf16. Until then, for Krea2 treat the result as "numerically equivalent", not "bit-identical".

A ComfyUI custom node pack. Version 0.1.0, experimental. Each module is import-safe: if it cannot load in your ComfyUI, it is skipped and the rest keeps working.

Modules

| module | what | status | |---|---|---| | layer | token-chunked execution of big diffusion transformers to lower activation memory (MiniMax H3 video, Krea 2 images). Nodes LayerStream: H3 Chunking, LayerStream: Krea2 Chunking | tested on one machine (RTX 4090, ComfyUI 0.39.0). Read its README for numbers and limits | | wos | small utility nodes. Megapixel Size: takes an input image and a megapixel budget and returns width, height (aspect ratio kept, rounded to a multiple of 8/16/32/64) and the real megapixels. Handy to set the video size from the start image | pure arithmetic, no model code. Unit-tested (tests/test_megapixel_size.py) |

Install

Copy this folder into ComfyUI/custom_nodes/ (folder name osiworx-pack) and restart ComfyUI. Nodes appear under osiworx/layer (search "LayerStream") and osiworx/image (search "Megapixel Size").

Megapixel Size and H3: use multiple = 32 for MiniMax H3 (32-pixel tokens). The node's default is 16, which is fine for most other video models but can give a size H3 does not divide evenly. 1.05 MP with 32 gives exactly 1024x1024 for a square image.

Test workflows (workflows/)

Copy to ComfyUI/user/default/workflows/ or drag into the UI. Pick your own model files in the loader nodes (the names are the test machine's).

• layerstream_krea2_lowvram.json: Krea2 (24.5 GB bf16), 2048x2048, chunking node + tiled VAE decode.
• layerstream_h3_15s_1mp.json: MiniMax H3 image-to-video, about 1 MP (a square start image gives 1024x1024; width/height come from the Megapixel Size node), 357 frames (about 15 s), 6 steps, turbo LoRA, with audio. Bypass the chunking node to test the no-node route. To pin the size by hand instead, delete the Megapixel Size node and type width/height into the H3 node.

Launch flags (scripts/)

The nodes handle activations; weights are streamed by ComfyUI's Dynamic VRAM. On a large card the flags are optional; to force a small budget steer it with:

--vram-headroom N --disable-pinned-memory (N = total VRAM minus the budget you want). scripts/run_lowvram.bat and scripts/run_h3_15s.bat are examples. Without --vram-headroom the card fills with weights and activations spill into shared system memory, which is very slow.

Honest summary of the first module

• Production report (one run, default launch flags, 24 GB card): Krea2 bf16 (24.5 GB) ran in about 1 minute with the Krea2 node and did not finish without it. Details and caveats in layer/README.md.
• H3 15 s at 1 MP: on the 24 GB test card the launch flag alone already fixed it; the node additionally keeps PyTorch's own allocations under a hard 8 GB cap with bit-identical output, about 10 % slower. We did not test cards below 24 GB; one user reported a 15 s clip finishing on a 12 GB card with 64 GB RAM (slow, about 35 minutes), see layer/README.md.
• Krea2: at 1 MP the flags are enough (no node needed). At 4 MP the node lowers the floor by roughly 1 GB (stock 3 GB, node 2 GB, with tiled VAE); output is numerically equivalent, not bit-identical, and 2.7-4x slower.
• Full details, limits and what was not tested: layer/README.md.

License

MIT, see LICENSE.

Layout

osiworx-pack/
  __init__.py          loads the modules
  pyproject.toml
  layer/               module: chunked execution (h3_chunking.py, krea2_chunking.py, README.md)
  wos/                 module: utility nodes (megapixel_size.py)
  tests/               plain-python tests (no ComfyUI needed)
  workflows/           test workflows
  scripts/             example launchers

Adding a module: create a folder with an __init__.py that exposes NODE_CLASS_MAPPINGS / NODE_DISPLAY_NAME_MAPPINGS, then add its name to MODULES in the top-level __init__.py.