Document measured encoder and runtime tradeoffs
Browse files
README.md
CHANGED
|
@@ -22,6 +22,24 @@ Idol itself accepts images, not text. A separately configured image generator ca
|
|
| 22 |
|
| 23 |
The original unmodified model card is included as [UPSTREAM_README.md](UPSTREAM_README.md); the original model, use conditions, and research are also described at [speridlabs/idol](https://huggingface.co/speridlabs/idol).
|
| 24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
## Evaluation limits
|
| 26 |
|
| 27 |
On one NVIDIA GeForce RTX 4090 workstation, with 50 sampling steps, 262,144 Gaussian points, original-precision image encoders, resident graphics-device models, batched guidance, and the experimental fused and optimized-attention paths, two image reconstructions were compared with original Idol weights. Relative latent error was 3.23% for the lamp and 6.47% for the axolotl. These are agreement measurements, not ground-truth 3D accuracy. In three fixed diagnostic orthographic views of the axolotl, calibrated Q8 had lower foreground color error than an otherwise comparable uncalibrated Q8 candidate; silhouette overlap improved in two of the three views.
|
|
|
|
| 22 |
|
| 23 |
The original unmodified model card is included as [UPSTREAM_README.md](UPSTREAM_README.md); the original model, use conditions, and research are also described at [speridlabs/idol](https://huggingface.co/speridlabs/idol).
|
| 24 |
|
| 25 |
+
## RTX 4090 observations
|
| 26 |
+
|
| 27 |
+
Individual Windows runs used an NVIDIA GeForce RTX 4090 with 24,564 MiB of physical graphics memory, PyTorch 2.10.0 with CUDA 12.8, and `triton-windows` 3.5.0.post21. Image-to-3D requests used 50 denoising steps, seed 42, 262,144 Gaussian points, and fused kernels. The resident rows also used batched guidance and optimized attention; staged rows disabled both. Warm requests followed a one-step warmup. Operating-system caches could be warm.
|
| 28 |
+
|
| 29 |
+
| Workload | Idol and image encoders | Time | Peak PyTorch GPU allocation | Peak host working set |
|
| 30 |
+
|---|---|---:|---:|---:|
|
| 31 |
+
| Fresh process, lamp / axolotl | Original Idol / original encoders | 37.0 / 36.0 s | 4.76 GB | ~6.43 GB |
|
| 32 |
+
| Fresh process, lamp / axolotl | Idol Q8 / original encoders | 34.6 / 34.6 s | 4.39 GB | ~6.73 GB |
|
| 33 |
+
| Warm resident, lamp / axolotl | Original Idol / original encoders | 25.12 / 25.12 s | 5.36 GB | ~6.07 GB |
|
| 34 |
+
| Warm resident, lamp / axolotl | Idol Q8 / original encoders | 28.05 / 27.38 s | 4.73 GB | ~6.72 GB |
|
| 35 |
+
| Warm staged, axolotl | Idol Q8 / original encoders | 33.28 s | 3.51 GB | 6.72 GB |
|
| 36 |
+
| Warm staged, axolotl | Idol Q8 / third-party DINOv3 Q4 and original FLUX.2 autoencoder | 30.92 s | 2.31 GB | 5.42 GB |
|
| 37 |
+
| Browser launch, prompt and result; two runs | Iris-3B Q4 and Qwen3-VL-4B Q4, then Idol Q8 / original encoders | 81.68 s; 85.08 s | Not measured end to end | Not measured end to end |
|
| 38 |
+
|
| 39 |
+
GB means one billion bytes; MiB means 1,048,576 bytes. PyTorch allocated memory excludes its unused reserved memory, the graphics context, the separately run Iris generator, and other processes. The table does not establish minimum required graphics memory or a distribution of latency. Image generation and Idol run sequentially, so their separate memory peaks must not simply be added. The image-model runner's configured 16 GiB limit is a budget, not a measured peak.
|
| 40 |
+
|
| 41 |
+
The third-party [DINOv3 Q4 file](https://huggingface.co/zisisnotzis/dinov3-gguf) was tested at revision `7fbb86b66bbb1dd483c94c179d35c969c76467e6`; its SHA-256 is `cf80ee6e3c2d4044b3214c0ccde5e416cb999b799f35b699f4edaa4059d83d71`. It is an experimental compatibility result, **not a quality recommendation**: for the axolotl image and matched staged settings, its latent differed by 34.17% in relative Euclidean norm from the same Idol Q8 run using the original DINOv3 encoder. One image and seed are insufficient to establish general behavior. The FLUX.2 autoencoder remained at original precision, and its separate terms apply.
|
| 42 |
+
|
| 43 |
## Evaluation limits
|
| 44 |
|
| 45 |
On one NVIDIA GeForce RTX 4090 workstation, with 50 sampling steps, 262,144 Gaussian points, original-precision image encoders, resident graphics-device models, batched guidance, and the experimental fused and optimized-attention paths, two image reconstructions were compared with original Idol weights. Relative latent error was 3.23% for the lamp and 6.47% for the axolotl. These are agreement measurements, not ground-truth 3D accuracy. In three fixed diagnostic orthographic views of the axolotl, calibrated Q8 had lower foreground color error than an otherwise comparable uncalibrated Q8 candidate; silhouette overlap improved in two of the three views.
|