Animation Generation

1. Install ComfyUI via Stability Matrix.
2. ComfyUI Manager → Install Models → search "animatediff" models (motion module + SD 1.5 base model).
3. AnimateDiff workflow:
   • Load SD 1.5 base model (4 GB) + AnimateDiff motion module (1.5 GB)
   • Prompt: "Animated landscape, clouds moving across sky, waterfall flowing, trees swaying gently, looping animation, 2D animation style, Studio Ghibli inspired."
   • Set frames=16 (1 second at 16 fps), context_length=16
   • Queue → first animated clip in 20-40 seconds on 8+ GB GPU
4. For longer animations: AnimateDiff supports 32-64 frames with context scheduling. 64 frames (~4 seconds) in 2-5 minutes.
5. For SDXL-based animation: AnimateDiff-XL (~8 GB total) at 30-60 seconds per 16-frame clip, higher quality.
6. Use cases: looping backgrounds, animated textures, motion graphics elements, character idle animations.

Used RTX 3060 12 GB ($200-250, see /hardware/rtx-3060-12gb). Runs AnimateDiff (SD 1.5) at 20-40 seconds per 16-frame clip (1 second of animation). For a 3-second looping animation: ~1-2 minutes. AnimateDiff-XL at 40-80 seconds per 16-frame clip. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe. Total: ~$390-440. Animation generation (AnimateDiff) is dramatically lighter than text-to-video — it's based on SD 1.5/SDXL, not full video diffusion models. At $400, you can generate looping animations in reasonable time.

Used RTX 3090 24 GB ($700-900, see /hardware/rtx-3090). Runs AnimateDiff-XL at 20-40 seconds per 16-frame clip, can handle 64-frame sequences (4 seconds) in 3-6 minutes without temporal artifacts. For production animation (game sprites, motion graphics, background loops), batch generation of 50-100 clips overnight is practical. Total: ~$1,800-2,200. For the fastest iteration: RTX 4090 ($2,000, see /hardware/rtx-4090) at 10-20 seconds per 16-frame clip. Animation generation is 10-50× faster than text-to-video.

The mistake: Setting frames=64 with context_length=16 on AnimateDiff and expecting a smooth 4-second animation — instead getting 4 distinct 1-second clips with jarring transitions every second. Why it fails: AnimateDiff's context_length is the temporal attention window — it only looks at 16 frames at a time. With 64 frames and context=16, frames 17-32 have no temporal relationship to frames 1-16. You get 4 independent 1-second generations concatenated. The fix: Use context scheduling (AnimateDiff's "context_stride" or "context_overlap" settings). This creates overlapping temporal windows: frames 1-16, frames 9-24, frames 17-32, etc. The overlap zones create smooth transitions. Or: use AnimateDiff-Lightning (4-step sampler) which handles longer sequences natively. For very long animations (100+ frames), generate in overlapping chunks and crossfade.