---
title: Krea Realtime Video
emoji: 🎬
colorFrom: blue
colorTo: purple
sdk: gradio
app_file: app.py
hardware: zerogpu
python_version: "3.10.13"
short_description: "Realtime T2V demo via diffusers ModularPipeline"
---

# Krea Realtime Video 14B

This Space implements **real local inference** for [krea/krea-realtime-video](https://huggingface.co/krea/krea-realtime-video) using the official 🧨 Diffusers `ModularPipeline` path from the model card.

## What it does

- **Text-to-video**: enter a prompt, choose how many blocks (≈ frames) to generate, and hit **Generate Video**.
- The model is a 14B autoregressive video diffusion model distilled from Wan 2.1 T2V 14B.

## How to try it

1. Open the **Generate Video** tab.
2. Type a prompt (e.g. *"a cat sitting on a boat"*).
3. Adjust **Number of Blocks** (1–9) and **Inference Steps per Block** (1–20).
   - Lower values = faster and cheaper first test.
4. Click **Generate Video**.

**Example inputs**
- `a cat sitting on a boat` (3 blocks, 4 steps, seed 42)
- `a futuristic city at sunset` (3 blocks, 4 steps, seed 123)
- `a panda playing guitar in a forest` (2 blocks, 4 steps, seed 7)

## Limitations & known risks

- **VRAM**: the 14B transformer (~28 GB in bf16) plus the UMT5-XXL text encoder (~12–14 GB) and VAE push the full pipeline well above 48 GB. Standard HF Space GPU tiers are **unlikely to fit** the model without aggressive quantization or >96 GB of VRAM.
- **Hardware recommendation**: this demo should be run on a Space with a very large GPU (e.g. `xlarge` ZeroGPU tier or a dedicated high-memory GPU).
- If the pipeline fails to load, the `/health` endpoint and the UI surface the **concrete load error** (OOM, CUDA init, missing import, etc.).
- The model card’s recommended optimizations (`torch.compile`, SageAttention, FP8 quantization with `torchao`) are **not enabled** here because:
  - `torch.compile` is [not supported on ZeroGPU](https://huggingface.co/blog/zerogpu-aoti).
  - `flash-attn` / `kernels` / `torchao` are native-build dependencies that frequently fail in HF Space build environments.
- We therefore rely on the **SDPA fallback** and vanilla bf16/fp16 precision.

## Architecture

- `app.py` loads `ModularPipeline.from_pretrained(..., trust_remote_code=True)` at module level so ZeroGPU can pack resident tensors.
- Generation is guarded by `@spaces.GPU` and uses iterative `PipelineState` blocks exactly as the model card demonstrates.
- Health (`api_name="health"`) is instant and never touches weights.

## Dependencies

- `diffusers` (main branch from GitHub)
- `torch`, `transformers`, `accelerate`, `safetensors`, `spaces`, `gradio`, `tqdm`