cstr/MI-GAN-512-GGUF

MI-GAN 512 (Places2) — GGUF for CrispEmbed

Experimental. These files run only in CrispEmbed's native inpainting engine behind default-off gates (CRISPEMBED_MIGAN=1, CRISPEMBED_MIGAN512=1, CRISPEMBED_INPAINT_QUANT=1 for F16/Q8_0/Q4_K). They are exact copies of the qualified files; nothing was requantized for publication.

Licence

• Code: MIT. Weights: MIT (LICENSE-WEIGHTS), Copyright (c) 2024 Picsart AI Research (PAIR).

Licence and notice files: licenses/LICENSE, licenses/LICENSE-WEIGHTS.

Lineage

• code: https://github.com/Picsart-AI-Research/MI-GAN/tree/2b793c5ece43f4253e32d4afc257120a5deed6f5
• code license: MIT
• weights license: MIT
• checkpoint: migan_512_places2.pt
• checkpoint sha256: 1d6087eee0aac8923ad2606be5d8caeb4824d3e4de331995e420c74e124a466a
• checkpoint sha256 status: first-observed (upstream does not declare a digest)
• Converter: models/convert-migan-to-gguf.py (via tools/ci-heavy/migan512_qualify.sh); quantizer: crispembed-quantize (inpaint-v1 policy).
• Publication source: private preservation repository at an exact revision (see manifest.json); CrispEmbed source 4290176e9eef708d7466708dad2d41afe6cd6b99.

Files and precision against the acceptance bars

Bars (final-output cosine, minimum over cases): F32 > 0.99, F16 > 0.98, Q8_0 > 0.97, Q4_K > 0.94. Synthetic-panel values compare with the original PyTorch model; real-photo values compare the masked region with native F32 (no original output exists for those photos).

Variant	File	Bytes	SHA-256	Min cosine	Bar	Notes
F32	migan-512-places2-f32.gguf	29500736	4bec681ef10f8ca914c5331f007a0605773cd09106d4fd6d9fcc0992f0a61739	1	0.99
F16	migan-512-places2-f16.gguf	17982208	932cfbd35db478dba4372f7ad9889ee68c84eac1c0b46025cf7336c8d55c0033	0.999998	0.98
Q8_0	migan-512-places2-q8_0.gguf	12583168	efdf1b08207f2c47c4710ca66831e6aff86ea31fca68593337f2ce605db70280	0.9965	0.97
Q4_K	migan-512-places2-q4_k.gguf	11143424	bd90218cd6f86a9b215fa3528ad5400536b7cf883a76e2450f22c835e376af46	0.977	0.94	0.9770 minimum on the synthetic panel; on the real-photo panel the street crop with a DBNet text mask reaches only 0.9019 masked-region cosine against native F32 (below 0.94). Exact-mask photos >= 0.962.

Caveat: Q4_K is 0.9019 on the street crop with a DBNet text mask (masked-region cosine against F32), below the 0.94 Q4_K bar on that crop; every other case clears it.

Use

See CrispEmbed docs/inpainting-cli.md and docs/inpainting-precision.md. Non-F32 files additionally need CRISPEMBED_INPAINT_QUANT=1.