Get started

For the original-FL / Larry 4+6 picture and twelve-step native-audio runtime on a complete 96 GiB SM120 GPU, see segmented AV. It is an explicit Python API with its own assets; H3Pipeline, catalog profiles, CLI and native HTTP below retain their separate contracts and defaults.

Vflash turns a prompt and optional images into a complete MP4 with audio. Most users want the complete pipeline (vflash generate, H3Pipeline or the Docker pipeline image). Applications that already build their own conditioning can use the lower-level native denoiser, which returns latents instead of video (below).

Nothing is downloaded or compiled implicitly: you choose a profile that matches your GPU, prepare its assets once, and Vflash refuses requests outside that profile's contract instead of silently changing the model, device or precision.

1. Install

Linux, Python 3.11+, an NVIDIA driver for CUDA 13.0, and ffmpeg/ffprobe on PATH. The base install is a light CLI without PyTorch; add the extra for the interface you use.

git clone --branch v0.22.1 --depth 1 https://github.com/Hansimov/vflash.git
cd vflash
python -m venv .venv && source .venv/bin/activate
python -m pip install -e '.[pipeline]'     # complete MP4 pipeline (includes GPU dependencies)

Use .[gpu] for the native denoiser only and .[server] for the native HTTP service. The Docker images pin the same environment (PyTorch 2.11 / CUDA 13.0) if you prefer containers.

The GitHub v0.22.1 release also provides a wheel and source archive from the same tag. Install the complete pipeline from the wheel:

python -m pip install "vflash[pipeline] @ https://github.com/Hansimov/vflash/releases/download/v0.22.1/vflash-0.22.1-py3-none-any.whl"

2. Check your GPU

vflash doctor      # GPUs visible to this process
vflash profiles    # profiles in this release

Pick the profile for your hardware and let plan check the match without loading weights:

GPU	Typical profile	Command
RTX 4090 48 GB	i2va-turbo4-v01-544-exact-sm89 (keyframes, hybrid, native HD)	vflash plan i2va-turbo4-v01-544-exact-sm89 --gpu 0
RTX PRO 6000 96 GB	i2va-turbo4-v01-544-exact-sm120 (hybrid image modes; see hardware limits)	vflash plan i2va-turbo4-v01-544-exact-sm120 --gpu 0
H100 SXM 80 GB	i2va-turbo4-v01-544-exact-sm90 (hybrid I2VA)	vflash plan i2va-turbo4-v01-544-exact-sm90 --gpu 0
Two RTX 3080 20 GB	ref2va-turbo4-exact-sm86	vflash plan ref2va-turbo4-exact-sm86 --gpu 0 --peer-gpu 1 --strategy sequence-head

Text-only (t2va-turbo4-exact-sm89), reference-image (ref2va-turbo4-exact-sm89) and Base16 keyframe profiles are also available on the 4090. The memory sizes are requirements: 24 GB 4090s and 10/12 GB 3080s are not supported configurations. See choose hardware for host RAM, residency and the measured envelope of each GPU.

3. Prepare the model assets

A profile needs the official MiniMax H3 components at a pinned revision, its pinned Turbo LoRA (if any), and three compiled native files (artifact, schedule and auxiliary tensors). Download and compile them with the official-weight compiler, describe the six paths in a JSON file (fields), then bind them once:

vflash prepare-pipeline --assets pipeline-assets.json --receipt prepared-assets.json

The receipt records the hash of every consumed file. Later starts only check identity and timestamps, so keep the assets in their final read-only location; moving them requires a new receipt.

4. Generate a first video

Write the prompt to a UTF-8 file and run one request. With a first-frame image (I2VA):

vflash generate --prepared-assets prepared-assets.json --prompt-file prompt.txt \
  --first-frame first.png --width 928 --height 512 --duration 5 --seed 1234 \
  --gpu 0 --output video.mp4 --trust-local-code

• Text only: prepare t2va-turbo4-exact-sm89 and omit the image arguments.
• Reference images: repeat --reference (refer to them as <Picture 1>, <Picture 2> … in the prompt).
• Last frame or both ends: --last-frame, or --first-frame plus --last-frame.

Progress is printed as JSON lines on stderr and the result on stdout. --trust-local-code allows the official decoder code from your verified local snapshot; review the licenses first. Each CLI call loads and releases the models, so the first run includes model loading.

Fifteen-second requests, up to nine references and native HD up to 2720×1536 use the hybrid model from Python; the generation guide shows how.

5. Keep the models loaded

For more than one request, keep one process alive:

• Python: one H3Pipeline serves requests serially and reuses loaded weights.
• Docker / HTTP: the native service queues jobs and keeps its model resident.

Measure the first (cold) request separately from repeated (warm) requests; see speed, memory and quality.

Native latents (advanced)

vflash denoise and the native HTTP service take a compiled conditioning bundle and write video and audio latents (.safetensors), not an MP4. Use them only if your application owns conditioning and decoding:

vflash denoise ref2va-turbo4-exact-sm89 --gpu 0 \
  --artifact /path/to/artifact --schedule-overlay /path/to/schedule \
  --auxiliary-tensor /path/to/auxiliary.safetensors \
  --bundle /path/to/bundle --output-latents result.safetensors

Two RTX 3080s cooperate on one request with --peer-gpu 1 --strategy sequence-head (the default strategy when a peer is selected; tensor is the alternative). Results need not be bitwise identical to single-GPU execution. The runtime assets page defines the four inputs.

If a check fails

Symptom	What to check
No GPU listed	nvidia-smi; in Docker, that the device is exposed to the container.
Profile does not match	Choose the profile for the compute capability and memory doctor reports; renaming a profile never converts assets.
Assets rejected	Re-run prepare-pipeline on the final, unchanged files.
Request rejected before loading	Canvas, duration or input count is outside the modes and limits.
Out of host memory	Block-ring and SM86 workers keep weights in RAM: allow at least 64 GiB available per worker.

More cases are in troubleshooting.