WAS/Pika-H3-Flex-PTD

Pika H3 Flex PTD

Pika H3 Flex merged into MiniMax H3's four-step PTD files at strength 0.5: Pika's motion and camera work, with sound, in four steps. Each file is one model patch for the H3 PTD Patch node of WAS Node Suite.

fl2va: first frame to video and sound	Ref2VA: one reference picture
pika_h3_flex_td_4step_s050 on minimax_h3_fl2va_int8_convrot. MiniMax H3 Image to Video starts on pikatd_lighthouse.png. Seed 4242.	pika_h3_flex_td_ref2va_4step_s050 on minimax_h3_ref2va_pruned_int8_convrot. MiniMax H3 Reference to Video takes pikatd_denim_portrait.png as <Picture 1>. Seed 42.
fl2va: text to video and sound, with speech	Text only, no picture
	An elderly bookseller in a hushed bookshop at night says: Every book in here has waited years for the right reader. Take your time. pika_h3_flex_td_4step_s050 on minimax_h3_fl2va_int8_convrot. MiniMax H3 Image to Video with no first frame. Seed 43.

File	Size	Transformer	Built from
pika_h3_flex_td_4step_s050.safetensors	4.79 GB	MiniMax H3 fl2va	Pika H3 Flex fl2va adapter + Nynxz's int8 PTD fl2va 4-step from nynxz/MiniMax-H3-PTD-LoRAs, made from epfl-vita/PTD-weights
pika_h3_flex_td_ref2va_4step_s050.safetensors	4.82 GB	MiniMax H3 Ref2VA	Pika H3 Flex ref2va adapter + Nynxz's int8 Ref2VA PTD 4-step from nynxz/MiniMax-H3-PTD-LoRAs, made from taobao-content-ai/MiniMax-H3-PTD-LoRAs

Use in ComfyUI

1. Put the file in ComfyUI/models/ptd or ComfyUI/models/model_patches.
2. Load the MiniMax H3 transformer the file is for: fl2va for the fl2va file, such as minimax_h3_fl2va_int8_convrot.safetensors, and ref2va for the ref2va file, such as minimax_h3_ref2va_pruned_int8_convrot.safetensors.
3. Pick the file in Load Model Patch and wire both into H3 PTD Patch.
4. Sample 4 steps at cfg 1: SamplerCustomAdvanced with euler and H3 PTD Patch's sigmas output, or KSampler with euler, simple, 4 steps.

No Pika node and no separate Pika LoRA are needed; the adapter is inside the file. Other LoRAs still apply on top. A sampler given sigmas off the file's trained steps stops with a message naming the right ones.

Example workflow

workflows/minimax-h3-pikatd.json renders the three clips above, two with the fl2va file and one with the Ref2VA file: 1344x768, 124 frames (5 seconds at 24 fps), 4 steps of euler on H3 PTD Patch's sigmas. It also ships with WAS Node Suite as docs/workflows/minimax-h3-pikatd.json. Put the two pictures from examples/ in ComfyUI/input.

Files the workflow loads

File	Folder	Download
minimax_h3_fl2va_int8_convrot.safetensors	ComfyUI/models/diffusion_models	Comfy-Org/MiniMax-H3
minimax_h3_ref2va_pruned_int8_convrot.safetensors	ComfyUI/models/diffusion_models	Comfy-Org/MiniMax-H3
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors	ComfyUI/models/text_encoders	Comfy-Org/MiniMax-H3
minimax_h3_video_vae_int8_convrot.safetensors	ComfyUI/models/vae	Comfy-Org/MiniMax-H3
minimax_h3_audio_vae_fp32.safetensors	ComfyUI/models/vae	Comfy-Org/MiniMax-H3
Both PikaTD files	ComfyUI/models/ptd	This repository

The workflow names the two transformers inside a MiniMax subfolder of diffusion_models; pick them again in Load Diffusion Model wherever they sit.

What is inside

Each file is one of Nynxz's int8 PTD four-step files unchanged, its per-step adapters and output heads, with the Pika H3 Flex adapter stored once as a shared. section and ptd_shared_strength 0.5. H3 PTD Patch applies the shared adapter at every step beside that step's own adapters.

File	Built from	Repository	SHA-256
pika_h3_flex_td_4step_s050	ptd/ptd_h3_4step_int8.safetensors	nynxz/MiniMax-H3-PTD-LoRAs	9e774393...
	comfy/loras/pika_h3_flex_fl2va_comfy_bf16.safetensors	Pika-Labs/Pika-H3-Flex	68783e64...
pika_h3_flex_td_ref2va_4step_s050	ptd/ptd_h3_ref2va_4step_int8.safetensors	nynxz/MiniMax-H3-PTD-LoRAs	f12f71ba...
	comfy/loras/pika_h3_flex_ref2va_comfy_bf16.safetensors	Pika-Labs/Pika-H3-Flex	e4d058fb...

Each file's ptd_shared metadata records both of its sources by name and full SHA-256.

Modified files

These files are modified versions of the sources above (MiniMax H3 Community License Agreement, Section III.2). Nynxz made the int8 PTD files from the EPFL VITA and taobao-content-ai releases, quantizing the adapter factors to int8 with one scale per row, and the NOTICE of nynxz/MiniMax-H3-PTD-LoRAs lists each change. WAS added a Pika H3 Flex adapter to each as a shared section at strength 0.5; the tensors of both sources are otherwise unchanged. Checksums are in SHA256SUMS.

Credits

Pika H3 Flex: Pika (Mellis, Inc.). Int8 PTD files: Nynxz, nynxz/MiniMax-H3-PTD-LoRAs, in the format of Nynxz's ComfyUI-PTD-Compact loader. PTD fl2va: EPFL VITA Lab, "Parametric Trajectory Distillation for Few-Step Video Generation". Ref2VA PTD: taobao-content-ai. Base model: MiniMax H3 by MiniMax.

License

MiniMax H3 Community License Agreement. See LICENSE and NOTICE.