Workflow library
Official templates, low-VRAM combos, speed distillations, long-clip sparse attention and batch recipes — each tells you what it needs, whether your machine runs it, and how long it takes.
Updated 2026-09-28
10 recipes
Official T2V template (text-to-video, 20-step baseline)
ComfyUI's built-in H3 text-to-video template: drop in your prompt and run. The 20-step baseline is the most stable quality-wise — best for first runs and quality-critical shots.
Official I2V template (image-to-video)
Use one image as the exact first frame and develop forward; composition and identity come from the image, so focus the prompt on camera and motion. See the I2VA mode in our prompt guide.
Official Ref2V template (multi-reference control)
Assign roles to images, video and audio (identity / style / motion / voice) and write the prompt in the six-section reference structure; best for character consistency and motion transfer.
8-step Turbo — the safe default (matches official turbo_mode)
Same as official turbo_mode: 20 steps down to 8 with minor quality loss — the least fiddly daily driver; see the live arena rank on its LoRA page.
Low-VRAM recipe: GGUF Q3 + 4-step (8–12GB cards)
The most realistic combo for 8GB-class cards: Q3 model (8.9GB) + Q2 encoder + 4-step distill; 540p short clips render. Full download stays under ~25GB.
FastH3 VSA 4-step (experimental, fastest)
FastVideo's 4-step VSA distill with NVIDIA (90% sparsity); biggest gains on long sequences, still an experimental path (live arena rank on its LoRA page).
PDD official acceleration, 8-step (dedicated node required)
Alibaba PAI's Parallel Decoding Distillation with a head bank; plain LoRA loaders silently drop it; live arena rank on its LoRA page.
Long-clip recipe: SLA sparse attention + SLA Turbo LoRA
The longer the sequence, the bigger the win: measured 44→31/25 s per step on a 5090 at 768p/15s (0.85/0.90 sparsity). Must be used as the SLA kit.
Parasyte Turbo 6 NFE (arena i2v #1)
Top bracket of the H3 Acceleration Arena (top-10 in both slices, highest vote count); 6 NFE ≈ 4.1× timed speedup — the current sweet spot.
Batch recipe: CLIPCached conditioning cache
Caches conditioning/latents to disk: from the second seed on, the whole text-encoding/conditioning stage is skipped (measured 950s→96s per clip on a 24GB card, source: pepikir).
Min VRAM / workable GPUs are estimated at 1344×768 · 5s with 32GB RAM and fast-disk. We host workflow files only with permission; others link to official or author sources.