I spent Thursday morning running the same Seedance 2.0 prompt through three different pipelines to see which one broke first. The answer was: none of them, and that's the story.
Here's what I saw this week.
1. Seedance 2.5 Goes Global — 3 Minutes of Cinema-Grade Video
ByteDance opened Seedance 2.5 early access outside China through Jimeng, and this is the week's biggest story by a mile. The jump from 2.0 is not incremental: 30 to 180 seconds of video in a single generation. That's three minutes of cinema-grade consistency without stitching ten-second clips together.
I watched @matchaman11's demo and the difference is visible in the first frame. The spatial coherence holds through camera pans, cuts, and scene transitions. If you're still planning shots in 10-second blocks, the math just changed. You can now plan scenes that last — not just flash cuts.
2. ComfyUI × Seedance 2.0: The Consistency Pipeline That Actually Works
@hellorob built a pipeline using DepthAnything V3 blended with Canny edge lines, gated through a SAM3 mask, and controlled entirely from a single ComfyUI MCP chat session. The result: consistent product shots where tiny debossed logos survive between takes. The whole thing is now a reusable Claude skill. I ran it. It works. This is the kind of pipeline that makes feature-length AI films possible — one character, one lighting setup, one chat session directing every shot.
3. Pika Director's Suite — Private Beta Changes the Game
@aimikoda tested Pika's Director's Suite in private beta and reported something that made me re-read the thread: zero regenerations. Feed it a plan, bring your own Anthropic key, and it orchestrates five parallel Seedance 2.0 generations — character sheets to finished cuts. The only edit was trimming one clip. The workflow is story → environment refs → parallel renders → auto-editable timeline. This is the video agent signal we've been waiting for.
4. GenBench Update: Grok 4.5 High Takes the Lead
Our GenBench leaderboard updated this week with fresh numbers. Grok 4.5 High hit 90.7 — that's +5.5 over GPT-5.5 High at a fraction of the cost ($3.40 vs $11.09). GPT-5.6 Terra is the best value in the OpenAI family: 87.9 average with a 91.7% episode score at half Sol cost. Full breakdown at genflick.com/genbench.
5. KlingAI's 4K Short Film Contest — Cannes Bronze
KlingAI took home a Cannes Lions Bronze and announced a 4K Short Film Contest. The bar for AI filmmaking is moving from "impressive for AI" to "impressive, period."
That's the week. Next time I'm digging into the Seedance 2.5 character consistency tests — the results are worth the wait.
GenFlick is building the AI filmmaking stack.
Sign up at genflick.com →