🎬 Full video tutorial:
Direct short movie scenes from one to three of your own pictures, with camera moves, dialogue, sound and music made together, locally, in six sampling steps. It uses Pika H3 Flex from Pika Labs on MiniMax H3 (open weights, not the pika.art app), plus the VSA speed gate and an optional Semantic Bridge switch.
In the graph: one group per reference picture, or First Frame Mode to start from a picture, a megapixels setting, a video with audio out, and a 2x RTX upscale output. Notes inside list every model link and the tested settings.
Tested settings (already set): sampler lcm, Pika's six manual sigmas, CFG 1, adapter strength 1, gate sparsity 0.75. Keep the memory cleanup nodes on.
Semantic Bridge: in my sitcom test the picture looked better with it on, and the sound was better with it off. Its author lists reference workflows as not supported in v1 and warns about reference-audio lip sync, so try both.
Measured on an RTX A4500 (20 GB), one reference, gate on:
5 s at 0.6 MP (1056x608): 1:43 sampling, 2:11 with the gate off
15 s at 0.6 MP: about 7:00 sampling, about 11:15 for the whole run with the 2x upscale
16 and 12 GB, simulated by capping VRAM on the 20 GB card, one reference: a 5 s clip at 0.6 MP stayed inside both caps at the same speed. A 15 s clip finished but peaked at about 17 GB, past both caps, so start with short clips on those cards.
0.8 MP at 15 s ran out of memory on 20 GB. 1344x768 at 15 s with three references ran out on 20 GB and finished on a 24 GB card.
Prompt tips:
Make each clip one complete little story in one location, then cut between clips.
Write positions the way they appear on screen ("the left of the frame"), and say where the camera is in every shot.
Write out why things happen. "The gap is too narrow for the drone's rotor arms" gave a crash you can actually follow.
Use things the model already knows. For something invented, give it a reference picture.
With several references, give each picture its own definition, and use only the appearance of a character sheet, without the panels.
Test your prompt on a short clip first.
Models:
minimax_h3_ref2va_pruned_int8_convrot.safetensors (about 21 GB), diffusion_models: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffusion_models
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (about 16 GB), text_encoders: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/text_encoders
minimax_h3_video_vae_fp16.safetensors (about 5.2 GB) + minimax_h3_audio_vae_fp32.safetensors (about 600 MB), vae: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae
pika_h3_flex_ref2va_v1_bf16.safetensors (about 2 GB), loras: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/loras
fasth3_vsa_gate.safetensors (about 3.9 GB), loras: https://huggingface.co/barelymining/ComfyUI-MiniMax-H3-FastVideo/tree/main
MiniMaxH3_SemanticBridge_v1.safetensors (about 11 MB), create models/semantic_bridge: https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge/tree/main
About 48 GB in total. Official model page: https://huggingface.co/Pika-Labs/Pika-H3-Flex
Custom nodes:
Ref2VA VSA by Kablex: https://github.com/Kablex/ComfyUI-Ref2VA-VSA
MiniMax H3 Semantic Bridge by speach1sdef178: https://github.com/Speach1sdef178/MiniMax-H3-Semantic-Bridge
Memory Cleanup, Nvidia RTX Nodes, KJNodes, rgthree, VideoHelperSuite: ComfyUI Manager, Install Missing Custom Nodes
Use PyTorch built for CUDA 13 (cu130) with an NVIDIA driver 580 or newer, and a recent ComfyUI (tested on 0.44.0).
Free guide with the same workflow: https://www.patreon.com/TheLocalLab/posts/174460405
Credits: Pika Labs (Pika H3 Flex), MiniMax (H3), the FastVideo team and Kablex (VSA gate), speach1sdef178 (Semantic Bridge).
The Pika H3 Flex adapters and MiniMax H3 are under the MiniMax H3 Community License Agreement. Read it before commercial use.