Document VP logSNR stability patch in SemDisDiffAE version history
Browse files
README.md
CHANGED
|
@@ -17,6 +17,7 @@ library_name: fcdm_diffae
|
|
| 17 |
|
| 18 |
| Date | Change |
|
| 19 |
|------|--------|
|
|
|
|
| 20 |
| 2026-04-10 | Refresh standalone package: fix bf16 RMSNorm precision path in both encoder and decoder to match training code; local export tooling now preserves fp32 EMA weights for future re-exports |
|
| 21 |
| 2026-04-08 | Fix posterior VP interpolation to use float32 precision (was using model dtype) |
|
| 22 |
| 2026-04-07 | Rename package `capacitor_diffae` → `fcdm_diffae`, class `FCDMDiffAE`; encode() now returns whitened latents, decode() dewhitens internally |
|
|
|
|
| 17 |
|
| 18 |
| Date | Change |
|
| 19 |
|------|--------|
|
| 20 |
+
| 2026-04-18 | Stabilize VP posterior and schedule alpha/sigma computation for extreme log-SNR values while keeping the public posterior interpolation path in float32 |
|
| 21 |
| 2026-04-10 | Refresh standalone package: fix bf16 RMSNorm precision path in both encoder and decoder to match training code; local export tooling now preserves fp32 EMA weights for future re-exports |
|
| 22 |
| 2026-04-08 | Fix posterior VP interpolation to use float32 precision (was using model dtype) |
|
| 23 |
| 2026-04-07 | Rename package `capacitor_diffae` → `fcdm_diffae`, class `FCDMDiffAE`; encode() now returns whitened latents, decode() dewhitens internally |
|