Pika H3 ComfyUI Setup: FL2VA, Ref2VA & VRAM
Pika H3 ships as open-weights adapters on top of MiniMax H3. Running it locally in ComfyUI gets you full parameter control, no per-second billing and no watermark. This page covers the install, the minimum workflow and the VRAM-vs-resolution trade-offs. Updated November 2026.
Install in 4 steps
- Update ComfyUI to the latest nightly (see docs.comfy.org).
- Download the FL2VA and Ref2VA safetensors from huggingface.co/Pika-Labs/Pika-H3-Flex.
- Place them in
ComfyUI/models/diffusion_models/with the original filenames. - Load the official workflow JSON from the same repo and queue a prompt.
VRAM & resolution
| Target | VRAM | Notes |
|---|---|---|
| 1080p / 10s | 24 GB | Single pass, ~7 seconds wall-time |
| 2K / 10s | 32 GB (or tiled) | Tiling costs 1.5–2x wall-time |
| 4K / 10s | 40 GB+ (or tiled) | Expect multi-pass refinement |
Minimum workflow
Five nodes: Checkpoint Loader → Adapter Loader (FL2VA or Ref2VA) → CLIP Text Encoder → KSampler (LCM, steps 6) → VAE Decoder. For Ref2VA, add the Reference Loader node between the Adapter Loader and the sampler.
FAQ
Which ComfyUI version is required?
Any recent nightly from docs.comfy.org. The Pika H3 nodes are included in the mainline branch as of November 2026, so no custom node pack is strictly required.
How much VRAM do I need?
24 GB is comfortable for 1080p at six steps. 2K needs 32 GB or tiled resampling; 4K needs 40 GB+ or multi-pass tiling. Below 16 GB the model loads with CPU offloading but output time climbs quickly.
Can I use both FL2VA and Ref2VA in one workflow?
Not in a single sampler call. The two adapters are loaded as alternate heads on the same base model; a chained workflow can call FL2VA for a draft and Ref2VA for a refinement pass.
Where do the weights go?
Under ComfyUI/models/diffusion_models/. Pull the FL2VA and Ref2VA safetensors from huggingface.co/Pika-Labs/Pika-H3-Flex and keep the original filenames.