anima
decoder
sorryhyun commited on
Commit
06ee509
·
verified ·
1 Parent(s): e3655ce

Card: benchmark table transposed, total time per image

Browse files
Files changed (1) hide show
  1. README.md +15 -10
README.md CHANGED
@@ -43,17 +43,22 @@ Outside those settings the head still decodes; quality there is unmeasured.
43
 
44
  ## Measured
45
 
46
- RTX 5070 Ti, 16 prompts at ~1 MP, compiled DiT blocks, one seed. PSNR is head vs the VAE
47
- decode of the same final latent.
 
48
 
49
- | sampler | DiT per image | head decode | VAE decode, resident / moved CPU↔GPU / loaded from disk | head vs VAE (median) |
50
- |---|---|---|---|---|
51
- | turbo, 4 steps, CFG 1 | 0.97 s | 7 ms | 141 / 176 / 851 ms | 35.4 dB |
52
- | base, Euler 28 + Spectrum, CFG 4 | 6.74 s | 14 ms | 141 / 179 / 796 ms | 33.0 dB |
53
-
54
- Peak VRAM of the decode above the resident DiT: 0.24 GB for the head, 1.30 – 1.38 GB for the
55
- VAE. On turbo the VAE swap is 15 % of DiT + decode time and the head under 1 %
56
- (about 1.17× images per second, text encoding excluded).
 
 
 
 
57
 
58
  Held-out reads during training (dB, colour / mono):
59
 
 
43
 
44
  ## Measured
45
 
46
+ RTX 5070 Ti, 16 prompts at ~1 MP, compiled DiT blocks, one seed; medians. PSNR is head vs the
47
+ VAE decode of the same final latent. "VAE, moved" is the usual interactive path: the VAE waits on
48
+ the CPU and goes to the card for the decode.
49
 
50
+ | | turbo, 4 steps, CFG 1 | base, Euler 28 + Spectrum, CFG 4 |
51
+ |---|---|---|
52
+ | DiT sampling | 0.97 s | 6.74 s |
53
+ | decode, pixel head | 7 ms | 14 ms |
54
+ | decode, VAE resident | 141 ms | 141 ms |
55
+ | decode, VAE moved CPU↔GPU | 176 ms | 179 ms |
56
+ | decode, VAE loaded from disk | 851 ms | 796 ms |
57
+ | decode peak VRAM above the DiT, head / VAE | 0.24 / 1.30 – 1.38 GB | 0.24 / 1.30 – 1.38 GB |
58
+ | head vs VAE | 35.4 dB | 33.0 dB |
59
+ | **total per image (DiT + decode), head vs VAE moved** | **0.98 s vs 1.15 s (1.17×)** | **6.75 s vs 6.92 s (1.02×)** |
60
+
61
+ Text encoding (~0.3 s per prompt) is outside the totals.
62
 
63
  Held-out reads during training (dB, colour / mono):
64