Hand off
Render settings for clay videos, depth passes and key stills, and what each video model accepts from Blender.
Render guide frames at the exact aspect ratio, frame rate and length of the clip you will generate, with no overlays and no depth of field. Then give each model only the inputs it accepts, within its limits. A clay video at 1280x720 and 24 fps meets the size and frame rate limits of every model here that takes a video. Veo takes stills only.
Render settings for guide frames
| Setting | Use | Why |
|---|---|---|
| Aspect and resolution | Match the target output exactly. 16:9: 1280x720 for the clay video, 1920x1080 for key stills. 9:16: 720x1280 or 1080x1920 | Seedance 2.5 needs at least 407,696 pixels per frame (854x480 passes, 640x360 fails), 300 to 6000 px per side, aspect 0.4 to 2.5. Kling Motion Control and Kling O3 edit need at least 720 px per side |
| Frame rate | 24 fps. Set it before you animate | Seedance accepts 24 to 60 fps. Wan 3.0 outputs 30 fps |
| Duration | Exactly the clip you will generate | Every model has an input cap (table below) |
| Engine | Workbench, or EEVEE with flat materials | See clay materials below |
| Overlays | None: no gizmos, grids, camera cones, motion paths, text or frame counters | ByteDance asks for this in the Seedance 2.5 guide |
| Depth of field and motion blur | Off | They also make the depth pass noisy |
| Output | PNG sequence with Overwrite off, then encode | A crash can resume. Blender 5.x can also write video directly (Media Type: Video) |
| Key stills | A full look render at the shot's exact first and last frames, or a clay frame restyled with an image model | For first and last frame and keyframe modes |
Encode the PNG sequence into an H.264 MP4 (yuv420p):
ffmpeg -framerate 24 -i clay/%04d.png -c:v libx264 -pix_fmt yuv420p -crf 16 03_clay.mp4Clay materials
- Coarse clay (most shots). Workbench with object colours is the fastest and gives each colour coded stand in a flat, distinct colour. EEVEE with flat emission materials works too.
- Fine clay (fully modelled scenes). EEVEE with grey materials and simple lighting. The video model then re-renders ("colours") the whole modelled scene.
Depth and other passes
- Depth. Use the Mist pass, not the raw Depth pass. Mist maps distance to 0 to 1; set start and depth in World, Mist Pass. Blender's Mist is dark near the camera, while most depth tools are fed maps that are bright near the camera, so invert it. Check one frame against the tool's own depth preprocessor before you render the full sequence. The inversion is our recommendation, not vendor guidance, so verify it for each workflow.
- Normals. Blender's Normal pass is in world space, and many control models expect camera space. Let the tool derive normals itself (for example Luma's
normalscontrol). - Pose. Blender does not render OpenPose skeletons. Render the rigged character as clay and let the tool's preprocessor extract the pose (OpenPose or DWPose in ComfyUI, Luma's
posecontrol).
Hand off per model
Prices checked on 15 Oct 2026.
| Model | Give it | Limits that matter |
|---|---|---|
| Seedance 2.5 | The clay video as @Video 1 (camera, motion, optionally lighting), colour mapped image references, keyframe images in order, or a first and last frame pair | Up to 10 videos (30 s total) and 30 images; mp4 or mov; 24 to 60 fps; 200 MB; no real faces. First and last frame mode and edit mode force an adaptive aspect ratio. Draft mode gives a 480p preview first |
| Kling Motion Control | A rendered character image plus a performance clip | Reference clip 3 to 30 s, at least 720 px, simple background, same aspect as the output. API 0.126 USD/s at 720p, 0.168 USD/s at 1080p |
| Kling 3.0 Omni and O3 | A clay or restyled video as the edit input or camera reference, or start and end frames | 3.0: one video (max 15 s) plus 4 images, or 7 images. O3 edit through Runway: up to 10 s, 720 to 3840 px, 4 image references. Kling 4.0 announces 10 keyframes and 5 videos (30 s), not confirmed shipped |
| Runway Aleph 2.0 | The clay video or a Blender render as input, plus up to 5 restyled keyframes pinned by seconds or by fraction of the clip | Input up to 30 s; optional outpaint to a new aspect ratio. API 0.28 USD/s |
| Luma Ray3.2 edit | The clay video as the source, with keyframes rendered from Blender | Source up to 18 s and 200 MB; up to 64 keyframes; adhere keeps structure; controls for pose, depth, normals, trajectory and face; output keeps the source aspect. For plain generation, up to 16 keyframes per clip |
| Wan 2.2 Fun Control (local ComfyUI) | A control video: inverted Mist depth, Canny edges, pose or trajectory | On an RTX 4090D, 81 frames at 640x640 took about 520 s (fp8) or 79 s (4 step LoRA) |
| Veo 3.1 | Two Blender stills: the first frame (image) and the last frame (lastFrame) | 4, 6 or 8 s; 16:9 or 9:16; 24 fps; 1080p and 4K need 8 s. No video reference, so the camera move must fit between the two frames |
Other models take Blender output too. Wan 3.0 on the API takes up to 5 reference videos (15 s total) or a first and last frame pair, but not both at once. MiniMax H3 takes up to 9 images and 3 videos (15 s total).
Which model for which output
- A clay video and character sheets: Seedance 2.5. ByteDance documents clay references officially for it.
- A finished Blender render to restyle: Runway Aleph 2.0 for the longest input with pinned keyframes, or Luma Ray3.2 edit when structure must hold.
- Body motion from a real performance: Kling Motion Control, with a still of your rendered character.
- Only a first and last frame: Veo 3.1 or Kling 3.0. Keep the change between the two frames small.
- A local GPU and no budget: Wan 2.2 Fun Control with the inverted depth pass.
Check limits before you render
Render to the strictest limit you will meet. A 640x360 clay video fails on Seedance, and a 15 s clip is too long for Kling O3 edit. Re-rendering a clay pass is cheap; a rejected upload after a long render is wasted time.