---
license: mit
library_name: crispembed
tags:
- gguf
- inpainting
- image-inpainting
- crispembed
---

# PixelHacker inpainting — GGUF for CrispEmbed

**Experimental.** These files run only in CrispEmbed's native inpainting engine behind default-off gates (`CRISPEMBED_PIXELHACKER=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

- Weights: MIT (author weight repository).
- Code (upstream PixelHacker): Apache-2.0. Native ports also study flash-linear-attention (MIT).

Licence and notice files: `licenses/pixelhacker-Apache-2.0.txt`, `licenses/pixelhacker-weights-model-card.md`, `licenses/flash-linear-attention-MIT.txt`, `licenses/NOTICE.md`, `licenses/SOURCE.txt`.

## Lineage

- code: https://github.com/hustvl/PixelHacker/tree/f5567db2871598aa178fe7a34c520dd478a0b41b
- code license: Apache-2.0
- weights: https://huggingface.co/hustvl/PixelHacker/resolve/012fd343158936a265b8a0ee38a791a7a2841f45/ft_places2/diffusion_pytorch_model.bin
- checkpoint sha256: ef5e28199cd03344ac22a6e6ff456ba8b868e83fb90d8174cc2ad83d7ef4296c
- weights license: MIT (author weight repository)
- 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 | `pixelhacker-f32.gguf` | 3449075872 | `5b486aa5d80647dbf298b0296d2cea7cb410dfb978b8bdfea806617f73ea588d` | 1 | 0.99 | pair with SDXL-VAE F32 |
| F16 | `pixelhacker-f16.gguf` | 1725648672 | `aaa3450a7c3640156b03f9cc0d5fda948ff6f94e144d06a9afe943b3023d8493` | 0.999999 | 0.98 | pair with SDXL-VAE F16; synthetic panel only; no F16 photo run |
| Q8_0 | `pixelhacker-q8_0.gguf` | 917886464 | `375c2b6ea621d17501b9693a12af6ac5b94ad1b17feb31821c5d4124fde3380f` | 0.9859 | 0.97 | pair with SDXL-VAE Q8_0 |
| Q4_K | `pixelhacker-q4_k.gguf` | 487079936 | `e34669015740fea59e9263e4ec89ee2569ed538d934a50661ae53b156cca5ae6` | 0.9425 | 0.94 | pair with SDXL-VAE Q4_K; borderline: 0.9425 on photos |

The SDXL VAE companion files are in [cstr/SDXL-VAE-inpaint-GGUF](https://huggingface.co/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`.