H3 Acceleration Options Compared: What to Pick, and What Stacks
H3 acceleration comes in three layers, and many tutorials lump them together — which is why you end up asking "why didn't stacking make it faster?" or "why did quality collapse?". Remember three categories first: step distillation (fewer sampling steps), attention backends (faster per-step compute), and caching (skipping part of the compute).
Layer 1: Step distillation (pick only one!)
| Method | Steps | Relative speed | Notes & dependencies |
|---|---|---|---|
| Official baseline | 20 | 1.0× | Most stable quality, slowest |
| LightX2V Turbo | 8 / 4 | ~2.1× / 3.4× | Built-in turbo_mode; 4 steps loses a little detail; the community favorite |
| larryvrh Turbo v4 | 8 | ~2.1× | Requires the dual-clock sampler (dual audio schedule); well regarded |
| FastH3 (FastVideo) | 4 | ~4× | DMD2 + VSA sparse distillation; on ComfyUI you need the converted LoRA |
| TaoMate-H3 | 3 | Fastest tier | Targets streaming long video; bigger quality cost |
| Alibaba PAI PDD | 8 / 4 | ~2× | Official distillation; must use the dedicated nodes and dedicated sigmas, and cannot stack with other distillation |
Layer 2: Attention backends (stackable)
- SageAttention: the default recommendation — 1.3-2× faster sampling, almost no quality loss. Note: 50-series cards risk silent noise on very long sequences (roughly 160k tokens or more); for long videos, switch to a sparse backend or Kitchen;
- comfy-kitchen built-in attention: bundled with ComfyUI 0.32+, same speed tier as Sage; pick it if you don't want to install anything;
- SLA sparse attention: requires the H3 SLA Attention node (Triton) used together with the SLA-trained Turbo LoRA as a set; measured on long videos at 44→25 s per step (5090 768p/15s), with limited gains on short clips;
- VSA (FastVideo Video Sparse Attention): the 90% sparse attention from the same source as the FastH3 4-step distillation; needs the VSA node, or enable it with the latest ComfyUI core Block Sparse Attention node (comfy-kitchen ≥0.2.33).
Layer 3: Caching (comes at a quality cost)
Step caches like EasyCache / Spectrum save another 15-40% of the time, but they work by "skipping part of the compute", so you need to spot-check frames; they conflict with PDD (they disrupt PDD's sigma boundaries). Conditional caching (e.g. CLIPCached) is lossless: when you generate several seeds in a row from the same prompt, from the second run on it skips the entire text-encoding stage.
Recommended combinations
| Your goal | Recommended combo |
|---|---|
| Fast idea scouting | LightX2V 4 steps + Sage (or Kitchen) |
| Balance of quality and speed | LightX2V 8 steps / larry 8 steps + Sage |
| Maximum speed (quality loss acceptable) | FastH3 4 steps + VSA (or Sage) |
| Official distillation route | PDD 8 steps + dedicated nodes (do not stack with any distillation/caching) |
| Batch output from the same prompt | Any combo above + CLIPCached conditional caching |
In step 2 of the calculator you can toggle these methods one by one: it automatically flags conflicts (for example, PDD can't be selected together with other distillation) and reflects the accelerated runtime in the results matrix. All the files live in the "Acceleration" category of the model library.