Seedance 2.5 Test 480p - 720p
I created a fight scene in a Chinese donghua style. For my prompt tests, I usually generate 5-second clips at 480p just to see how the style and overall look turn out.
Interestingly, action scenes look cleaner at 480p. At 720p, they become noticeably blurrier or noisier, even though both were generated with the exact same prompt.
The timing also seems much more consistent across generations now. Seedance 2.0 wasn't following timestamps reliably, so I rarely used them back then. With 2.5, I have started using timestamps again.
I upscaled the 480p with Topaz.
What do you think? Also you can check the prompt below.
AI won’t replace 3D artists by generating better characters.
It’ll replace the 11 hours spent stopping a skirt from passing through someone’s thigh.
Cloth like this looks effortless, but underneath is collision setup, weight painting, simulation, material tuning, and dozens of tiny fixes before the fabric moves naturally instead of exploding through the body.
That boring middle is exactly where models like Kimi K3 become dangerous: inspect the animation, detect clipping, adjust the Blender scene, rerun the simulation, and keep iterating.
Designing the character is the fun part. Fixing the same dress for the 37th time is the job AI will take first.
Seedance 2.5 frozen time + rewind effect prompt is below 👇
Photorealistic cinematic 1950s American diner, chrome stools, red vinyl, neon glow and checkerboard floor, shot with modern lived-in realism and soft natural window light. Subtle handheld texture, warm practicals, rich period detail, heavy film grain.
0-4s: [Medium Wide] A striking young woman in her early 20s sits alone at the counter, calm and slightly amused, slowly sipping a tall thick milkshake through a straw. Behind her a young waitress in classic uniform approaches with a tray of eggs and bacon in one hand and a full glass coffee pot in the other. An older lady starts rising from a nearby booth.
4-8s: [Dynamic Tracking] The older lady collides hard into the waitress. Tray, plate, eggs, bacon and coffee pot explode upward in chaotic slow motion. Coffee erupts into long liquid ribbons and perfect suspended droplets. Camera immediately begins a smooth continuous orbit around the impact. Time locks completely at the peak of the spill. Every face freezes in pure shock. Only the girl at the counter keeps moving, completely unfazed.
8-17s: [Slow 360° Orbital] Camera glides in a full elegant orbit through the frozen diner. Coffee hangs in mid-air as glassy ribbons and spheres with perfect volume and surface tension. Bacon strips, eggs and the spinning tray float weightlessly. Patrons and waitress remain locked in startled expressions. The girl at the counter takes one slow, deliberate sip, eyes half-lidded, almost bored, while the entire frozen world (except her) begins an elegant reverse: every droplet, every piece of food and every person rewinds smoothly back to the exact starting positions.
17-24s: [Medium Shot] Rewind lands perfectly. Waitress stands balanced again with tray and coffee pot. The girl lifts her eyes, raises two fingers in a small casual gesture and softly calls the waitress by name. The waitress turns toward her just before the older lady begins to stand, completely avoiding the collision. A tiny private smile crosses the girl’s face.
24-30s: [Extreme Close-Up] Hard cut to her face as she takes one last slow sip. Soft knowing smile, eyes almost closed in quiet satisfaction, like she has done this a hundred times. Shallow depth of field, creamy bokeh of the neon diner behind her.
Photorealistic, ultra-detailed fluid physics, perfect motion blur only on moving elements, stable characters, cinematic lighting, heavy natural film grain, no artifacts, movie-level temporal coherence, high rewatch value.
30 seconds long is the new AI standard! hope you like it 🫡
あれ? #MiniMaxH3 でMV作るのめちゃくちゃ楽しいかもしれない☺️
H3は耳も良いみたいですが、画像の文字やレイヤーみたいなのもしっかり認識してくれてるっぽいですね。
作り方
① Codexにキャラ画像と歌詞を渡して各歌詞付きの画像を一気に生成
② 曲のmp3を15秒ごとに分割
③ H3で画像とmp3ファイルをリファレンスさせて生成
リプ欄にイメージ付けておきますので、よろしければ参考にしてみてください☺️
@Hailuo_AI
BLENDER MCP + KIMI K3 JUST DID THIS
Top is a basic gray-shaded scene inside Blender. Bottom is the final cinematic result.
> same car
> same driver
> same camera path
> completely different level of realism and motion
Kimi K3 controlled the scene through Blender MCP.
The gap between simple 3D and final shot keeps getting shorter.
the results of my experiments are exceeding my expectations
I am spending a lot of money but it is all in the name of science. I have a good story in mind, I should try to turn it into a miniseries
터치만 하면 애니메이션으로! 마법의 펜 프롬프트
이런 감각적인 분위기, CF에서 좀 본 것 같지?
이 프롬프트 쓰면 너도 바로 딸깍 할 수 있어..!
준비물
1. Seedance 2.0이나 Mini에 넣고 10초로 제작
2. 변주를 주고 싶다면 GPT나 클로드랑 대화해봐!
프롬프트는 하단 게시글에!