Comparison summary
Use each model's tables for measured values; no universal ranking is implied.
Before reading the tables: see the labels and memory glossary for runtime labels, denoise/decode scope, tracked totals and GPU/RAM peaks.
Windows BF16 comparison
Existing single-run observations on an RTX 3060 Ti 8 GiB / 64 GiB RAM. Method names below identify the transformer; all rows in these primary Windows sets use a shared official leaf text encoder and resident VAE (temporal streaming for video). Component choices can be combined; a native transformer does not mean every pipeline component uses native VMM.
Times are seconds. GPU peaks are sampled absolute driver usage, RAM peaks are process RSS, both in GiB, over the tracked run. Tracked total excludes command startup and downloads. Denoise and other phase intervals can be nested; do not sum them to reconstruct total. See metric definitions.
VAE decode is the recorded decoder execution phase, excluding VAE loading. For video it means the video decoder only; audio decode and vocoder are separate phases in the detailed tables. Tracked total already includes decoding and other pipeline work. Memory columns remain global run peaks, not decoder-only peaks.
LTX video — 1280×704, 5 seconds, 121 frames, 8 steps
All three rows use the default PyTorch SDPA path for video self-attention. The external FlashAttention 2 and Sol-Attn options were not selected. The specific kernel chosen internally by SDPA was not recorded.
| Runtime method | Denoise (s) | Video VAE decode (s) | Tracked total (s) | Peak GPU (absolute, GiB) | Peak RAM (RSS, GiB) |
|---|---|---|---|---|---|
| Diffusers streamed leaf | 392.91 | 22.52 | 452.33 | 4.72 | 41.34 |
| DDO linear | 262.22 | 23.47 | 336.44 | 4.72 | 47.87 |
| DDO native VMM | 207.40 | 24.00 | 277.09 | 5.21 | 32.65 |
See each video phase and final total, including conditioning and decoder costs. This workload generates audio without input audio; later audio-conditioned recording recipes are different comparisons.
LTX image — 1280×704, 8 steps
| Runtime method | Denoise (s) | VAE decode (s) | Tracked total (s) | Peak GPU (absolute, GiB) | Peak RAM (RSS, GiB) |
|---|---|---|---|---|---|
| Diffusers streamed leaf | 39.03 | 0.32 | 69.97 | 7.15 | 28.77 |
| DDO linear | 11.99 | 0.30 | 54.33 | 7.43 | 27.12 |
| DDO native VMM | 15.15 | 0.39 | 48.73 | 7.27 | 29.53 |
Linear has shorter denoise; native has shorter total. Diffusers leaf has the lowest global GPU peak of these three rows. See image phases and final total.
Qwen Image 2.1 — 512×512, 20 steps
| Runtime method | Denoise/readback (s) | VAE decode (s) | Tracked total (s) | Peak GPU (absolute, GiB) | Peak RAM (RSS, GiB) |
|---|---|---|---|---|---|
| Diffusers streamed leaf | 41.28 | 0.47 | 57.50 | 3.68 | 18.97 |
| DDO linear | 17.94 | 0.49 | 41.36 | 6.69 | 21.58 |
| DDO native VMM — reused observation | 17.67 | 0.47 | 36.64 | 6.12 | 18.97 |
The native Qwen row comes from a separate session, not a side-by-side control; its small denoise difference from linear does not establish a reliable advantage. Diffusers leaf uses less GPU memory. See Qwen phases and final total and the separate 1024-square results there; resolutions must not be mixed.
These configured pipeline results do not isolate the effect of VMM, establish
an advantage over a matched torch.compile setup, or certify other models.
Native Linux, WSL, lower-RAM, VAE-route and quantization comparisons remain in
their separate groups. Hardware and environment limits.
Choose what to investigate
| Model / priority | What the selected measurements support | Tables |
|---|---|---|
| LTX image / one-shot speed | Windows linear has shorter denoise but native has shorter tracked total; Linux native is shortest in both primary measures. | LTX image |
| LTX video / one-shot speed | Native fast VMM has the shortest tracked total in each primary five-second platform set. | LTX video |
| Qwen / one-shot speed | Linear/native shorten 512-square denoise versus official leaf; native has the shortest tracked total in the selected sets. | Qwen |
| Lower transformer-phase process RAM | Windows LTX video protected NVMe uses much less denoise RSS with a close total to native; OS file cache and the encoder still matter. | Video Windows |
| Lower decoder GPU pressure | Leaf VAE can lower decode GPU peaks at extra time/RAM; use tiling if activations still do not fit. | Image VAE, video VAE |
| Saved INT4 | Selected Windows SDNQ DDO image cases outperform their INT4 leaf controls; this is not BF16 quality equivalence or native VMM support. | LTX INT4, Qwen INT4 |
| Saved FP8 | Selected LTX video pairs reduce process RSS but take longer and have higher global GPU peaks than BF16. | Video precision pairs |
| WSL | Bounded native or explicit-cap linear can trade time for GPU usage; Windows/Linux registration limits and streamed leaf performance do not transfer automatically. | Image WSL, video WSL, Qwen WSL |
These are single workload observations, not causal allocator-only or low-RAM hardware guarantees. Text encoding can dominate global RAM even when denoise RAM falls. Compare phase peaks and total separately; scenario links inside each table provide the relevant limitations.
See metric definitions, presets and reproduction.