Model assets 0.2.0

Model assets for Pickapicka's AI fill and generative fill. Not a code release.

Both are downloaded by the app only when the feature is first used, after it
says so, and each file is checked against the SHA-256 below.

migan_pipeline_v2.onnx (AI fill)

MI-GAN 512 (Places2), the authors' own ONNX pipeline, mirrored unchanged.

• Upstream: https://huggingface.co/andraniksargsyan/migan (Picsart AI Research, "MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices", ICCV 2023)
• Licence: MIT for the code and the weights (LICENSE and LICENSE-WEIGHTS in https://github.com/Picsart-AI-Research/MI-GAN)
• SHA256: 6f1f3530a1a2324b19752018ce756088b07973cda8d7d890034ace5c8a48c40b

Generative fill (sd15-inpaint-lcm, 10 files, 1.9 GB)

Stable Diffusion 1.5 inpainting with the LCM-LoRA for SD 1.5 fused in, as ONNX
for onnxruntime on the CPU. These files are under CreativeML OpenRAIL-M and
CreativeML OpenRAIL++-M, not MIT. Both allow commercial use and both carry
use-based restrictions (Attachment A) that bind anyone who uses or passes on
the files; the licences are included, and NOTICE.txt lists what was changed.
The app shows the restrictions and asks for them to be accepted before it
downloads these files.

• Upstream: https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-inpainting (CreativeML OpenRAIL-M) and https://huggingface.co/latent-consistency/lcm-lora-sdv1-5 (OpenRAIL++)
• Built by packaging/models/build_genfill.py (torch 2.14.1, diffusers 0.41.0, transformers 5.19.0, peft 0.21.2, onnx 1.23.2, onnxruntime 1.30.0): the LoRA fused at scale 1.0; the UNet, the VAE encoder (latent mean) and decoder exported to ONNX opset 17; the UNet's attention fused by onnxruntime's transformer optimizer with the CUDA-only fusions off; float weights stored as float16 with a cast to float32; the empty prompt's text embedding as empty_prompt.npy in place of the text encoder. A clean rebuild reproduced every model file byte for byte.
• Inputs: UNet sample [1, 9, 64, 64] float32, timestep [1] int64, encoder_hidden_states [1, 77, 768] float32; VAE encoder image [1, 3, 512, 512] in [-1, 1]; VAE decoder latent [1, 4, 64, 64] (divided by 0.18215).

File	Bytes	SHA256
unet.onnx	532093	aa18b01246f9f1ba3311e61d78a6aa71d7a92d9e9c709eedaf1038e494cb0e5f
unet.onnx.data	1719446400	4a3792122198dc04dead85e7e4bfd60d2cddf6fe29d8b548a1b78945fd640483
vae_encoder.onnx	122884	5998e90bf01a113bf8ee65d27d18239a6e7e4720995db3aade28469f58dfb5ff
vae_encoder.onnx.data	68375296	0941bad47aa4f5819076306be15aa8558bde8d50b0c4ff95bd4252cac1d2c000
vae_decoder.onnx	132127	26d86e74cfdedf4d74a21e6010c442020545cdc7c66bc59df9ffcf73045a0381
vae_decoder.onnx.data	99046144	726bc0bc3e4d8adc5300c81149821562f9c3bebcc5c6a3261965a4e3d9f2e430
empty_prompt.npy	236672	d2d2f2583ceded75236c4190bd1df1086d7ff7e608ca402210f21231c193b28b
LICENSE-CreativeML-OpenRAIL-M.txt	14385	be351ebe7ac01bcdbb018639aadcfd38f136b7dc3f2a3d4d3a24db51d1b210ef
LICENSE-CreativeML-OpenRAIL++-M.txt	14206	e514164be7f2ca2c0fb39e7283b01ab4a06d802837eaf171aa4645cfd7ee1157
NOTICE.txt	1597	9e0b27c1092045c21889722776ba2eeba7c37297e1aa7a7e7719c9845c5deb81