Deepmind1/Dream-wan2-2-fp8da-aoti-preview-2

Dream Motion Pro — Wan 2.2

Create cinematic videos from a single image or guide the animation with both a first frame and a last frame.

Main features

• —Single-image Image-to-Video mode
• —First + Last Frame mode
• —Wan 2.2 14B Lightning/MotionBoost checkpoint
• —FP8 transformer quantization
• —AOTI-compiled transformer blocks for ZeroGPU
• —Fast, Balanced and High Quality profiles
• —Natural, Cinematic, Portrait, Product Ad, Dynamic and Anime motion styles
• —Identity-preservation prompt controls
• —Automatic professional prompt enhancement
• —Landscape, portrait, square and social-media aspect ratios
• —Optional RIFE interpolation to 32 or 64 FPS
• —Seed control and repeatable generations
• —Extract any generated video frame and use it as the next first image

Space variables

Add these in Settings → Variables and secrets:

Recommended defaults

• —Quality Mode: Balanced
• —Duration: 3.5 seconds
• —Output FPS: 16
• —Steps: 6
• —Guidance: 1.0 / 1.0
• —Scheduler: UniPCMultistep
• —Flow Shift: 3.0

Lightning checkpoints usually perform best at approximately 4–8 inference steps. Raising the steps or guidance unnecessarily increases GPU time and may reduce motion quality.

First + Last Frame mode

Both images are normalized to the same dimensions. The last image is center-cropped to match the prepared first frame so the pipeline receives a consistent resolution.

This release intentionally supports one image or first + last frames only. It does not include a three-frame mode.

Notes

• —32 and 64 FPS outputs use RIFE after the base 16 FPS generation.
• —Higher FPS improves playback smoothness but does not create additional semantic motion.
• —A clear, high-resolution source image and one main action generally produce the most stable result.
• —Generate multiple seeds when choosing a final result.