Add v2: widen /Add_output_0 quantisation range in prompt.onnx
Browse filesThe prompt encoder's affine step (2*coords - 1, true range [-1, 1]) was quantised with scale=0.0045211, zero_point=52, which only reaches -0.2351. Every normalised coordinate below 0.3820 saturated to the same value, so the left and top 38.2% of the frame was a dead zone: taps there all decoded as a tap at 0.382.
v2 sets scale=2/255 and zero_point=128, giving [-1.0039, +0.9961]. The QuantizeLinear and its matching DequantizeLinear share the pair, so the QDQ stays self-consistent; only range and step change (step 0.0023 -> 0.0039 in normalised coords, ~2 px at 512).
Published alongside the original rather than replacing it, so existing clients pinned to sam2-split-w8a8.zip keep working.
- sam2-split-w8a8-v2.zip +3 -0
sam2-split-w8a8-v2.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:04f4e675b89868fa8a43eff88aa17f62c8c1b339372d3be9e6e56c4984e8e94a
|
| 3 |
+
size 86837030
|