seedance-vs-veo（种子 vs veo）

Getting Started

Paste a video prompt and I'll generate one MP4 from Seedance and one from Veo for side-by-side review. No prompt? Describe the scene you want to test.

Try saying:

• "Run 'a red sports car drifting on a rain-slicked road at night' through both Seedance and Veo and show me the MP4s"
• "Compare how Seedance and Veo handle the prompt 'a hummingbird hovering near a flower in slow motion' — I need to see motion quality differences"
• "Test this prompt on both models: 'a chef plating a dish in a modern kitchen' — include generation time and resolution in the results"

Automatic Setup

On first interaction, connect to the processing API before doing anything else. Show a brief status like "Setting things up...".

Token: If NEMO_TOKEN environment variable is already set, use it and skip to Session below.

Free token: Generate a UUID as client identifier, then POST to https://mega-api-prod.nemovideo.ai/api/auth/anonymous-token with header X-Client-Id: <uuid>. The response field data.token becomes your NEMO_TOKEN (100 credits, 7-day expiry).

Session: POST to https://mega-api-prod.nemovideo.ai/api/tasks/me/with-session/nemo_agent with Bearer auth and body {"task_name":"project"}. Save session_id from the response.

Confirm to the user you're connected and ready. Don't print tokens or raw JSON.

Run One Prompt, Get Two AI Videos Back

Type a single text prompt — say, "a cyclist riding through a foggy forest at dawn" — and the skill sends it to both Seedance and Veo simultaneously. You get 2 MP4 files, each labeled with model name, resolution, and generation time in seconds.

The comparison covers motion consistency, text adherence, and artifact frequency across a 3–8 second clip range. It doesn't rewrite your prompt for either model. What you write is what both models receive, keeping the test fair.

Results land in a side-by-side layout with metadata attached. You can re-run the same prompt 3 times to check output variance per model.

Routing Prompts Between Models

Your input gets parsed for model-specific keywords — 'Seedance' routes to ByteDance's API endpoint, 'Veo' routes to Google DeepMind's endpoint, and anything ambiguous triggers a side-by-side comparison job that fires both simultaneously.

API Calls and GPU Queues

Each generation request hits the respective cloud rendering pipeline — Seedance queues on ByteDance's distributed GPU cluster while Veo processes through Google's TPU-backed infrastructure, so wait times differ based on each provider's current load. Response payloads return a signed video URL, generation latency in milliseconds, and the exact prompt echo used after any internal rewriting.

Headers are derived from this file's YAML frontmatter. X-Skill-Source is seedance-vs-veo, X-Skill-Version comes from the version field, and X-Skill-Platform is detected from the install path (~/.clawhub/ = clawhub, ~/.cursor/skills/ = cursor, otherwise unknown).

All requests must include: Authorization: Bearer <NEMO_TOKEN>, X-Skill-Source, X-Skill-Version, X-Skill-Platform. Missing attribution headers will cause export to fail with 402.

API base: https://mega-api-prod.nemovideo.ai

Create session: POST /api/tasks/me/with-session/nemo_agent — body {"task_name":"project","language":"<lang>"} — returns task_id, session_id.

Send message (SSE): POST /run_sse — body {"app_name":"nemo_agent","user_id":"me","session_id":"<sid>","new_message":{"parts":[{"text":"<msg>"}]}} with Accept: text/event-stream. Max timeout: 15 minutes.

Upload: POST /api/upload-video/nemo_agent/me/<sid> — file: multipart -F "files=@/path", or URL: {"urls":["<url>"],"source_type":"url"}

Credits: GET /api/credits/balance/simple — returns available, frozen, total

Session state: GET /api/state/nemo_agent/me/<sid>/latest — key fields: data.state.draft, data.state.video_infos, data.state.generated_media

Export (free, no credits): POST /api/render/proxy/lambda — body {"id":"render_<ts>","sessionId":"<sid>","draft":<json>,"output":{"format":"mp4","quality":"high"}}. Poll GET /api/render/proxy/lambda/<id> every 30s until status = completed. Download URL at output.url.

Supported formats: mp4, mov, avi, webm, mkv, jpg, png, gif, webp, mp3, wav, m4a, aac.

Reading the SSE Stream

Text events go straight to the user (after GUI translation). Tool calls stay internal. Heartbeats and empty data: lines mean the backend is still working — show "⏳ Still working..." every 2 minutes.

About 30% of edit operations close the stream without any text. When that happens, poll /api/state to confirm the timeline changed, then tell the user what was updated.

Backend Response Translation

The backend assumes a GUI exists. Translate these into API actions:

Draft field mapping: t=tracks, tt=track type (0=video, 1=audio, 7=text), sg=segments, d=duration(ms), m=metadata.

Timeline (3 tracks): 1. Video: city timelapse (0-10s) 2. BGM: Lo-fi (0-10s, 35%) 3. Title: "Urban Dreams" (0-3s)

Error Codes

• 0 — success, continue normally
• 1001 — token expired or invalid; re-acquire via /api/auth/anonymous-token
• 1002 — session not found; create a new one
• 2001 — out of credits; anonymous users get a registration link with ?bind=<id>, registered users top up
• 4001 — unsupported file type; show accepted formats
• 4002 — file too large; suggest compressing or trimming
• 400 — missing X-Client-Id; generate one and retry
• 402 — free plan export blocked; not a credit issue, subscription tier
• 429 — rate limited; wait 30s and retry once

Quick Start Guide

Write a prompt between 10 and 80 words. Shorter prompts under 15 words tend to produce more variance between models, which makes the comparison more informative.

Submit it here — don't adjust phrasing for either model. The skill sends the identical string to both Seedance and Veo, then returns 2 labeled MP4 files, each capped at 1080p and between 3–8 seconds long.

Check the metadata block attached to each file. It lists model name, resolution, aspect ratio, and wall-clock generation time in seconds. Use those numbers when writing up your evaluation, not subjective impressions.

Re-run the same prompt 3 times if you need variance data. Output consistency differs between the two models, and a single run won't show that.

Troubleshooting

If one MP4 returns blank or corrupted, the model timed out — generation cutoff is 90 seconds per model. Resubmit the prompt once before assuming a model-side failure.

Prompts over 150 characters sometimes cause Veo to truncate scene elements. Split long prompts into a core scene description (under 100 characters) plus a style note to keep both models working from equivalent inputs.

If both clips look identical, your prompt is likely too generic. Add at least 2 specific visual details — lighting condition, subject motion, camera angle — to create testable differences between the models' outputs.

Resolution mismatches (e.g., one file at 720p and one at 1080p) aren't a bug. The two models don't share a resolution default. Check the metadata block on each file rather than eyeballing clip size.