Video Rebirth's compressed, few-step version of MiniMax-H3 (14B). Generates video and synchronized stereo audio together in 4–8 steps, no CFG.
Source: official Video Rebirth showcase. Clips include H3-generated stereo audio — unmute to listen.
HyperFlow H3 Flash has no hosted API yet, and no free tier can run its 14B weights (needs 24 GB+ VRAM plus a 15.7 GB text encoder). The practical free route is the MiniMax H3 family on hosted platforms. All of these run in the cloud — only a browser is required.
5 generations per day, up to 15 s each, 768p, native stereo audio. H3 Max is the flagship quality model (Design Arena image-to-video #1, Elo 1341) — higher quality than HyperFlow Flash. Free quota is sandbox-only.
Generate free on fal.aiAbout 2 clips/day at 20 credits per generation. Audio included. Daily credits reset automatically.
Open Google FlowAbout 2–3 short clips/day. No native stereo audio like H3.
Open PixVerseStacking the free tiers above gives roughly 9–10 free clips per day across platforms. Want HyperFlow H3 Flash specifically on an API? Vote for provider support.
Recommended default: Q4_K_M (9.1 GB). Needs the 15.7 GB qwen3vl_32b_minimax_h3_nvfp4_awq text encoder plus video/audio VAEs from Comfy-Org/MiniMax-H3.
# 1. clone custom nodes
git clone https://github.com/city96/ComfyUI-GGUF \
ComfyUI/custom_nodes/ComfyUI-GGUF
git clone https://github.com/Video-Rebirth/ComfyUI-HyperFlow-Flash \
ComfyUI/custom_nodes/ComfyUI-HyperFlow-Flash
# 2. put the .gguf in models/diffusion_models/
# 3. Workflow > Browse Templates > ComfyUI-HyperFlow-Flash
# templates: hyperflow_t2v, hyperflow_i2v,
# hyperflow_flf2v, hyperflow_ref2vFull multimodal references (images, videos, audio). First run downloads ~90 GB (~68 GB with fp8 text encoder).
pip install git+https://github.com/Video-Rebirth/hyperflow-h3-flash
hyperflow-h3-flash \
--prompt "A chef slides a steaming bowl of ramen; rain taps on the window." \
--steps 6 --output ramen.mp4
# keyframes: --image first.png --last-image last.png
# references: --reference person.png --reference scene.pngsd-cli -M vid_gen \
--diffusion-model minimax_h3_fl2va_hyperflow_flash_14b-Q4_K_M.gguf \
--llm qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors \
--vae minimax_h3_video_vae_fp16.safetensors \
--audio-vae minimax_h3_audio_vae_fp32.safetensors \
--sampling-method euler --cfg-scale 1.0 --steps 6 \
-W 1344 -H 768 --video-frames 124 --fps 24 -s 42 \
-p "A chef slides a steaming bowl of ramen; rain taps on the window." \
-o output.webmRequires the hyperflow-two-time.patch on upstream commit 3be2378. Add --offload-to-cpu on smaller GPUs.
| Task | Input | Notes |
|---|---|---|
t2va | Text prompt | Text → video + audio |
fl2va | Prompt + first keyframe (± last) | Image → video + audio |
ref2va | Prompt + up to 12 references | 9 images, 3 videos, 3 audio (GGUF: images only) |
Describe subject, action, setting, camera movement, and sound in one prompt.
No negative prompt (CFG = 1.0). For ref2va, cite inputs in order as
<Picture 1>, <Video 1>, <Audio 1>.
Steps: 4 (speed) · 6 (default) · 8 (quality).