OpenSourceVideoGen — Local AI Video Skill · Open Skill 公開技能

OpenSourceVideoGen — Local AI Video Skill

OpenSourceVideoGen — Local AI Video Skill 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。

Anime

15 1 11

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3762060/2026-09-27/8dyjpd32.jpg

# OpenSourceVideoGen — Local AI Video Skill

## When to use
- The user wants to make short AI videos for free using open-source models (text-to-video or image-to-video), running locally on their own GPU.
- The user asks which open-source video model to use, what hardware/VRAM they need, how to install ComfyUI or Diffusers, or wants ready-to-use prompt templates.
- The user wants free cloud fallbacks because they have no suitable GPU.
- The user wants a local Python helper or ComfyUI workflow to run inference.

This skill always prioritizes fully open weights with permissive licenses (Apache 2.0 / MIT). It never suggests paid closed APIs or credit systems as the primary method.

## Steps
1. **Clarify the setup** (if not already given): GPU VRAM (8/12/16/24 GB), ComfyUI vs pure Python (Diffusers), text-to-video (T2V) vs image-to-video (I2V) priority, and whether they need audio in the video. If unknown, ask before giving exact settings.
2. **Pick the default backend**:
   - No audio needed → **Wan 2.2 / 2.7** (Apache 2.0, best quality-per-VRAM, strong I2V identity consistency + motion, camera-motion LoRAs and ControlNet support). Default.
   - Audio needed → **LTX-2.5** (Apache 2.0, the only mainstream open-source model with built-in audio generation, fast).
   - HunyuanVideo 1.5 is higher quality but heavy — only recommend for 24 GB+ or heavy quantization.
3. **Give the model shortlist** (I2V-first), with license and VRAM:
   - Wan 2.2 / 2.7 (I2V) ⭐ — Apache 2.0; 1.3B ~8 GB, 14B ~24 GB (or ~12–14 GB GGUF Q4). Default.
   - LTX-2.5 — Apache 2.0; ~8–12 GB; fastest; built-in audio.
   - HunyuanVideo 1.5 — MIT weights; ~24 GB / ~16 GB quantized; top quality, heavy.
   - CogVideoX 5B — Apache 2.0; ~16–24 GB; good balance, strong I2V.
   - Mochi 1 — Apache 2.0; ~24 GB; strong motion, heavier.
   - Open-Sora — Apache 2.0; ~16–24 GB; research-friendly, flexible.
4. **Give hardware / quantization guidance** by VRAM tier:
   - 12 GB: Wan 1.3B I2V, Wan 14B GGUF Q3, LTX-2.5, CogVideoX 5B (fp8/GGUF).
   - 16 GB: Wan 14B I2V GGUF Q4_K_M ⭐, CogVideoX 5B (full), HunyuanVideo (quantized).
   - 24 GB: Wan 14B I2V (full), HunyuanVideo (full), Mochi 1.
   - Quantization: GGUF (ComfyUI-GGUF node) or fp8 (bitsandbytes) cuts VRAM ~30–50% with minor quality loss. Always enable `enable_model_cpu_offload()` + `enable_vae_slicing()` / `enable_vae_tiling()`.
5. **Provide the exact tuned setup** when VRAM is known. Reference setup for **16 GB VRAM, Wan 2.2 14B I2V GGUF Q4_K_M**:
   - Model: `Wan-AI/Wan2.2-I2V-14B-GGUF` (Q4_K_M or Q5_K_M); VRAM ~12–14 GB.
   - Resolution 832×480 (Wan native), 81 frames (~3.4 s @ 24 fps), 30 steps, CFG 6.0, euler/simple, Wan VAE (fp16), offload + vae_slicing + vae_tiling.
   - Why not 1.3B: 14B has better motion/detail/identity; at Q4 GGUF it still fits 16 GB. Why not full fp16: needs ~24 GB.
6. **Give step-by-step ComfyUI install** (see Notes for the exact commands) and the ready-to-copy workflow node graph: Load Image → Load Diffusion Model (Wan 14B I2V GGUF Q4 via ComfyUI-GGUF) → CLIP Text Encode (pos/neg) → KSampler (steps 30, cfg 6, euler/simple) → VAE Decode (Wan VAE fp16) → Video Combine (fps 24) → Save Video (MP4 H.264). For T2V, swap the model for the T2V GGUF and drop Load Image.
7. **Provide the Diffusers Python helper** (I2V-first, Wan 14B GGUF default) — see Notes for the full single-file script. Key points: `--image` is the default mode (I2V), `--no-image` falls back to T2V, `--backend` picks wan/ltx/hunyuan/cogvideox, and it includes VRAM-management + fp8/GGUF quantization comments.
8. **Give ready-to-use prompt templates** (positive + negative). For I2V the prompt is an animation instruction (motion, camera, lighting) while the input image carries the subject. See Notes for the template library.
9. **Offer free cloud fallbacks only when the user has no GPU**: Hugging Face Spaces (free CPU/limited GPU tiers), Google Colab (free T4, session-limited), Kaggle Notebooks (free ~30 h/week GPU), or self-hosted on a c

分享 Digen

OpenSourceVideoGen — Local AI Video Skill · Open Skill 公開技能

OpenSourceVideoGen — Local AI Video Skill