Add experimental IC-Light LCM-LoRA models
Browse files- lcm-sd15-v1/README.md +26 -0
- lcm-sd15-v1/iclight_unet_lcm.onnx +3 -0
- lcm-sd15-v1/iclight_unet_lcm.onnx.data +3 -0
- lcm-sd15-v1/manifest.json +14 -0
- .gitattributes +1 -0
lcm-sd15-v1/README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# IC-Light LCM-LoRA experimental UNet
|
| 2 |
+
|
| 3 |
+
Independent experimental weights for IC-Light Android 1.11 or later.
|
| 4 |
+
Original models in the repository root are unchanged. This subfolder contains
|
| 5 |
+
the fused UNet only; retain the original CLIP, RMBG and VAE models.
|
| 6 |
+
|
| 7 |
+
Use Settings → Download and enable LCM in the companion app. It downloads the
|
| 8 |
+
two ONNX files, resumes interrupted downloads, checks pinned SHA-256 hashes,
|
| 9 |
+
and selects LCM 4 steps / CFG=1. The model is loaded when generation starts.
|
| 10 |
+
The same two files can also be imported from a local folder. No ZIP extraction
|
| 11 |
+
is needed for automatic downloading.
|
| 12 |
+
|
| 13 |
+
Adapter: https://huggingface.co/latent-consistency/lcm-lora-sdv1-5
|
| 14 |
+
Adapter scale: 1.0. LCM scheduler: epsilon prediction, original inference steps
|
| 15 |
+
50, scaled-linear beta 0.00085–0.012, direction-conditioned strength 0.9.
|
| 16 |
+
The eight-channel IC-Light foreground input is preserved. Graph topology and
|
| 17 |
+
I/O match the deployed original ONNX; 278 parameter tensors changed.
|
| 18 |
+
|
| 19 |
+
Local portrait and food tests completed with 4 and 6 steps. Food images showed
|
| 20 |
+
stronger saturation and contrast; quality is not equivalent to the original.
|
| 21 |
+
Mac timings are not Android NPU timings. No LCM Qualcomm device acceptance test
|
| 22 |
+
has been completed. Keep the original 8-step ODE path for comparisons.
|
| 23 |
+
|
| 24 |
+
See manifest.json for exact file sizes and SHA-256. Source licenses for the
|
| 25 |
+
base model, IC-Light, and the LCM-LoRA adapter continue to apply; this experiment
|
| 26 |
+
does not relicense the derived weights.
|
lcm-sd15-v1/iclight_unet_lcm.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e8142e67b34b28c63571654261588bb8050cb75c3acc766353f42bcf7b786601
|
| 3 |
+
size 924921
|
lcm-sd15-v1/iclight_unet_lcm.onnx.data
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:12eb58018a7e8bb5dd1c04f6071e3f5fea92cfea0a64b8db998084717b089aee
|
| 3 |
+
size 1718999040
|
lcm-sd15-v1/manifest.json
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"sampler": "LCM",
|
| 3 |
+
"changed_tensors": 278,
|
| 4 |
+
"files": {
|
| 5 |
+
"iclight_unet_lcm.onnx.data": {
|
| 6 |
+
"bytes": 1718999040,
|
| 7 |
+
"sha256": "12eb58018a7e8bb5dd1c04f6071e3f5fea92cfea0a64b8db998084717b089aee"
|
| 8 |
+
},
|
| 9 |
+
"iclight_unet_lcm.onnx": {
|
| 10 |
+
"bytes": 924921,
|
| 11 |
+
"sha256": "e8142e67b34b28c63571654261588bb8050cb75c3acc766353f42bcf7b786601"
|
| 12 |
+
}
|
| 13 |
+
}
|
| 14 |
+
}
|
.gitattributes
CHANGED
|
@@ -38,3 +38,4 @@ iclight_unet_fc.onnx.data filter=lfs diff=lfs merge=lfs -text
|
|
| 38 |
iclight_vae_dec.onnx.data filter=lfs diff=lfs merge=lfs -text
|
| 39 |
iclight_vae_enc.onnx.data filter=lfs diff=lfs merge=lfs -text
|
| 40 |
rmbg.onnx.data filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 38 |
iclight_vae_dec.onnx.data filter=lfs diff=lfs merge=lfs -text
|
| 39 |
iclight_vae_enc.onnx.data filter=lfs diff=lfs merge=lfs -text
|
| 40 |
rmbg.onnx.data filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
[[:space:]]lcm-sd15-v1/iclight_unet_lcm.onnx.data filter=lfs diff=lfs merge=lfs -text
|