cstr/Moebius-inpaint-GGUF

Moebius inpainting — GGUF for CrispEmbed

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

Licence

• Code (upstream Moebius): Apache-2.0.
• Weights: the official Moebius README states Apache-2.0; the author's Hugging Face model card states MIT. Both records are reproduced verbatim in licenses/. No dual licence is claimed; the conversion only reproduces the upstream weights and their licence records.

Licence and notice files: licenses/moebius-Apache-2.0.txt, licenses/moebius-upstream-README.md, licenses/moebius-weights-model-card.md, licenses/NOTICE.md, licenses/AUTHOR-WEIGHTS-README.md, licenses/SOURCE.txt.

Lineage

• code: https://github.com/hustvl/Moebius/tree/b88d462bacb9af6e7128a3b4cc4a07418bedfd61
• code license: Apache-2.0
• weights: https://huggingface.co/hustvl/Moebius/resolve/cd01f47fb648219d3fa605806c5ce00b713faa5e/ft_places2/diffusion_pytorch_model.bin
• checkpoint sha256: 6525afb888e55f9b5c74fa0a5d19ca0762d720d6c716fb0f8422fbeb6868a09a
• weights license: Apache-2.0 (official README); the author HF card says MIT. Both records are preserved; no dual licence is claimed.
• Converter: models/convert-latent-inpaint-to-gguf.py; 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	moebius-f32.gguf	904877728	f7365937fb82665e5725ab78b701e9650316023d7291d48da22cd9ea0ec7e352	1	0.99	pair with SDXL-VAE F32
F16	moebius-f16.gguf	484668672	5ac92700dc65b6f6abe6c4d63c56172b9212eea34cf76b4b4a4898b9c1291750	0.999935	0.98	pair with SDXL-VAE F16
Q8_0	moebius-q8_0.gguf	287696832	2b5fc98e455d0df8240a823b5e025eee67b7a8dbb6c4067a08c3ed7b0b9f4de4	0.9795	0.97	pair with SDXL-VAE Q8_0
Q4_K-mix	moebius-q4_k-mix.gguf	246888384	4fcb546bf3bc4161bb91d33329b6ca0757872c66747d11414a2bc26e068bb0fe	0.9449	0.94	pair with SDXL-VAE Q8_0; Q4_K with every attention-block tensor at Q8_0; crispembed-quantize with CRISPEMBED_QUANTIZE_INPAINT_MIX=1 (tools/ci-heavy/inpaint_quant_policy_check.sh checks the bytes); 0.9931 synthetic, 0.9449 photos (dequantized arm; 0.9504 stored matmul)

Not published: Q4_K (plain Q4_K fails the bar on photos (0.859)).

The SDXL VAE companion files are in cstr/SDXL-VAE-inpaint-GGUF.

Use

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