mohsenvand/mindview-painter

mindview painter

The painter half of mindview, packed for WebGPU in the browser. It is what neovand.github.io/mindview downloads. These are not new weights for general use; they are a browser layout of existing ones, plus two small parts trained for mindview.

File	What it holds
manifest.json	Every tensor: shape, file and byte offsets; the model config and the default 4-step schedule.
dit_0.bin … dit_4.bin, dit_misc.bin	The diffusion transformer of Bonsai Image 4B (FLUX.2 [klein] 4B with ternary weights). Ternary tensors are 16 two-bit codes per u32 plus one f32 scale per 128 inputs; the rest is dense.
adapter.bin	A ternary linear adapter trained for mindview: it maps Ternary Bonsai 1.7B's states at layers 7, 14 and 21 to the DiT's 7,680-wide text conditioning, in place of the original text encoder.
taef2.bin	TAEF2, the tiny latent decoder.
viz.bin, viz.json	A per-block lens fitted for mindview (a linear readout of the picture the DiT has in mind after each block) and colour probes for the decoder's stages. For display only.
schedules.bin, schedules.json	Noise levels and the timestep modulation for other step counts.
export_report.json	How the export was checked.

The files are written by research/scripts/export_painter.py (with export_painter_schedules.py and export_painter_viz2.py) in the mindview repository. The browser runtime that reads them matches the PyTorch reference block for block.

Sources and licenses

• Bonsai Image 4B (ternary DiT) by PrismML, Apache 2.0. It is FLUX.2 [klein] 4B by Black Forest Labs, Apache 2.0, with its weights made ternary. The DiT here is repacked from prism-ml/bonsai-image-ternary-4B-unpacked@30bbc03 with every ternary weight exact (export_report.json). The timestep embedding is not included: its outputs, the modulation for each step of the schedules, are precomputed (mod.*), checked against float64.
• TAEF2 by Ollin Boer Bohan (madebyollin/taef2), MIT license; repacked, unchanged.
• The adapter and the lens were trained for mindview on the models above. They are released under Apache 2.0.
• The reader, Ternary Bonsai 1.7B (PrismML, Apache 2.0), is not included: the site downloads it from PrismML's repository.