用 AI 做视频,最头疼的不是慢,是没法改:一处不满意,整条重新抽卡。
Viggle 开源的 Remake 正好补上这块。
视频是官方 demo:《爱乐之城》同一个镜头,改出 22 个版本。
它是 MiniMax H3 上的两个 LoRA,把「重抽」变成「定向改」:
· edit:删掉多余的物体、加个道具、换背景 / 天气 / 光线
· swap:只换人或角色
· 动作、运镜、节奏、原声,全部沿用原片
用法:截一帧,用图像模型改成想要的样子,钉回原来的位置重渲;提示词里写明哪些保留,它照字面保留。
懒得自己画帧,就把视频和一句话丢给 Claude Code,切镜头、画帧、渲染它自己来。
门槛:官方测试表里 16GB 显存用 int8 能跑 832×480,跟我跑 H3 用的是同一种 int8 布局。
huggingface.co/Viggle/Viggle…
你现在一条 AI 视频,平均要抽几次才定稿?
试过了,在多视图这个场景下
P-H3-image 效果还不如 Qwen Image 2.1 好
完整的测试报告:
eternityspring.github.io/ai-…
开源图片模型都见过
蒸馏 MiniMax H3 这种视频模型而造出的开源图片模型: P-H3-Image 还是第一次见,ComfyUI 原生支持
不知道这种开源模型做场景多视图效果怎么样,等我先帮各位尝尝咸淡
huggingface.co/PrunaAI/P-H3-…
开源图片模型都见过
蒸馏 MiniMax H3 这种视频模型而造出的开源图片模型: P-H3-Image 还是第一次见,ComfyUI 原生支持
不知道这种开源模型做场景多视图效果怎么样,等我先帮各位尝尝咸淡
huggingface.co/PrunaAI/P-H3-…
Today, we open-source P-H3-Image, an image generation and editing model built from MiniMax-H3 by @MiniMax_AI. A SOTA video model, turned into an efficient open image model.
- 𝗛𝗼𝘄 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗶𝘀 𝗶𝘁? Generate or edit images 5× faster, in 8 steps instead of 50. A 4 MP image takes 2.6 s instead of 12.4 s. The model is 7× smaller than H3, so high-resolution images are a light download and run on a much smaller GPU.
- 𝗛𝗼𝘄 𝗴𝗼𝗼𝗱 𝗶𝘀 𝗶𝘁? Strongest at photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text. It renders 2 to 16 MP and edits with up to 8 input images.
-𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? We reworked every component of H3: the prompt upsampler, the text encoder, the diffusion schedule, the transformer, and the decoder. Use it with Diffusers or ComfyUI.
📄 Tech Report: pruna.ai/blog/oss-p-h3-image…
🤗 @huggingface Open Weights: huggingface.co/PrunaAI/P-H3-…
👉 @ComfyUI Native Support: huggingface.co/PrunaAI/P-H3-…
这下爽了,FastVideo 也支持 Pika-H3-Flex了!
不知道这双加速能碰出什么火花,据官方说最快提速 7倍
github.com/KadoBOT/ComfyUI-H…
Drama 3 本来是要至少 24G 显存才能跑
有人把他搬进 Comfyui 中,实测 16G 的4080 效果也不错
据作者说 8G 显卡就能跑,项目地址:
github.com/PedroMarinhoDev/C…
之前我基于 Drama 3 和 IndexTTS 做了评测,完整报告:
eternityspring.github.io/ai-…
这个 4B 模型有意思啊
画面 → 3D 场景,本地就能跑
效果谈不上炸裂,但拿来做 AI 视频运镜预演感觉足够了
重点是:自部署没成本
huggingface.co/ccore/Qwen3.5…
已交 30 块钱保护费,准备开整。看看多久能回本。
还是网友勤奋啊!整理了1407 张适合短剧服饰的穿搭图
能按关键词搜,每张只有衣服、没有人物,2:3 竖图ai-costume-gallery.pages.dev…
起初我也想做这种库,思路是整理提示词
看到它才反应过来:图片模型的参考图能力上来以后,一张图比一段提示词准得多,也省事得多
我准备做个类似的微信小程序,搜集整理更多品类的这种服饰穿搭图。
MiniMax H3 Frame Break Effect 测试提示词:
subject_definitions:
<Picture 1> is a six-panel storyboard sheet that plans the single continuous [Shot 1]. Its panels are read left to right, top to bottom; each panel is one moment of the same locked-off camera, with its time printed in the lower-left corner: 00:00, 00:01.5, 00:03, 00:04.5, 00:06 and 00:08. It defines the viewpoint, the letterbox composition, the subject placement and the order and timing of the beats. The white gutters between panels and the printed timecodes are planning marks only.
<Subject 1> is Taotao in <Picture 1>, a small young girl painter with messy peach-pink hair in two round buns, a red beret, thin round black-framed glasses, big brown eyes, a white T-shirt with pushed-up sleeves under baggy blue denim overalls splattered with cyan, yellow, red and purple paint, rolled cuffs, white socks and red canvas sneakers.
<Subject 2> is Niangao in <Picture 1>, a round, chubby white cat with orange patches on its back and head, a long white tail and a lazy, unbothered half-closed-eyed face.
<Subject 3> is the attic painter's studio in <Picture 1>: a sloped wooden ceiling with a skylight on the left, mint-green walls, a light wooden floor, an easel holding a colorful abstract painting, a bookshelf, potted plants, hanging pendant lamps and wooden worktables crowded with paint cans and brushes, plus one small cyan paint can.
<Subject 4> is the letterbox composition in <Picture 1>: a thin solid black bar across the top of the 16:9 frame and a thick solid black bar across the bottom, with the studio picture spanning the full width between them and no side borders.
summary:
[reference generation] The target video is one continuous, locked-off 10-second wide shot that plays out the six panels of <Picture 1> in order and on their printed times. <Subject 1> paints in the air in <Subject 3> while <Subject 2> sits on the lower bar of <Subject 4>; the cat's tail knocks the cyan paint can toward the lens, <Subject 1> dives to catch it, the leftover paint splats over the left of the picture, she presses her face to the lens with a wink, and she ends sitting on the lower bar with <Subject 2> in her lap beside a heart drawn in the paint, giving a thumbs-up.
retention_analysis:
<Picture 1> (storyboard for [Shot 1]): fully_preserved - the eye-level viewpoint, letterbox framing, subject placement, and the order and timing of all six panels are followed; the panel grid, white gutters and printed timecodes are not rendered.
<Subject 1> (appears in [Shot 1]): fully_preserved - Taotao's peach-pink double buns, red beret, round glasses, white T-shirt, paint-stained denim overalls and red sneakers stay identical throughout.
<Subject 2> (appears in [Shot 1]): fully_preserved - the chubby white cat with orange patches and its lazy expression are retained.
<Subject 3> (appears in [Shot 1]): fully_preserved - the attic studio with its skylight, easel, shelves, plants, lamps and paint-covered worktables is retained.
<Subject 4> (appears in [Shot 1]): fully_preserved - the two black bars keep the same position and thickness for the whole video.
detailed_description:
The target video is clean 2D cel-shaded anime with bold crisp ink outlines, flat color fills, hard-edged two-tone shadows and bright, playful colors, animated with fast, snappy comedy timing, quick anticipation and springy overshoot, while the cat stays lazy and slow. It is one locked-off wide shot at eye level with no cuts, no zoom and no camera movement. The two bars of <Subject 4> never move, bend, crack or flicker; anything that crosses the lower bar (the cat's tail, the paint can, paint drips, hands, arms, legs) is drawn in front of the black for a pop-out illusion, and nothing ever crosses the upper bar. There are no subtitles, panel borders, gutters or timecodes in the frame.
[Shot 1] The shot opens on the first panel of <Picture 1>. In <Subject 3>, <Subject 1> (S1) balances on one leg on a small wooden stool in the middle of the room, humming a cheerful wordless tune as she swings a big brush and paints a bright cyan arc in the air. On the right, <Subject 2> sits on the edge of the lower bar, its long white tail dangling in front of the black, next to the small cyan paint can. At 00:01.500, matching the second panel, the cat's tail whips across the can with a quick swish; the can tips off the bar and tumbles toward the lens, huge and foreshortened in front of the lower bar, cyan paint splashing out of it, while <Subject 1> gasps on the stool with her arms flung wide, mouth open and buns bouncing. At 00:03.000, matching the third panel, she dives forward off the stool toward the lens; her right arm stretches out, strongly foreshortened, and her hand snaps the can out of the air just in front of the lower bar with a clink, while the rest of her body stays inside the studio. At 00:04.500, matching the fourth panel, the leftover paint hits the lens with a wet splat: a big cyan splash covers the left third of the picture and drips run down over the lower bar. <Subject 1> sits frozen on the floor behind it, holding the can in both hands, a cyan dot on her nose and her mouth a small round "o". At 00:06.000, matching the fifth panel, she wipes her sleeve across the paint with a squeak and pushes her face right up to the lens in an extreme close-up, her forearm pressed over the lower bar, glasses knocked askew, flashing a toothy grin and a wink. She then hops up onto the lower bar with a soft thump, scoops <Subject 2> into her lap and quickly draws a heart in the cyan paint with the brush, which squeaks on the glass. At 00:08.000, matching the sixth panel, she sits on the lower bar with her legs and red sneakers dangling in front of the black and <Subject 2> curled contentedly in her lap, the heart in the paint beside her. The cat gives a small meow, and <Subject 1> winks at the lens and gives a thumbs-up, holding the pose until the end.
overall_soundscape:
Soft attic room tone continues throughout, under the humming, the tail swish, the can's clink and tumble, the wet splat, the squeaky sleeve wipe, the soft thump, the brush squeaking on glass and the small meow. There is no dialogue.
non_diegetic_music:
N/A
Frame Break Effect Demo 测试提示词:
请生成一张分镜故事板:3 行 2 列共 6 个画格,每格 16:9,格间留细白缝,整体白色底。每格左下角只用小字标注时间码,不要出现其他任何文字。
【6 格中必须完全一致的部分】
- 构图:每格上下各有一条粗的纯黑横条(各占画面高度约 15%),中间是画面区域。黑条的位置、粗细、形状在 6 格里完全一样,像印在屏幕上的平整色块,不能有裂纹、弯曲、破碎,不要竖向边框或画框。
- 画风:赛璐璐日系动画,粗而干净的墨线描边、平涂色块、硬边两阶阴影、无渐变,明亮俏皮;角色脚下有一小块柔和的地面阴影,体现纵深。
- 场景:画家阁楼工作室,平涂薄荷绿墙面,浅木色地板,角落有画架、几个颜料桶和一支大画笔。
- 角色桃桃:桃粉色双丸子头,红色贝雷帽,圆框眼镜,白 T 恤外面套沾满彩色颜料渍的蓝色牛仔背带裤,袖子卷到手肘,红色帆布鞋。6 格里脸型、发色、服装完全一致。
- 角色年糕:一只圆滚滚的白猫,背上有一块橘色斑,脸上一副懒洋洋的表情。
- 镜头:固定机位,平视,不推拉。
【6 个画格,从左到右、从上到下】
1. 00:00,宽景。桃桃很小,站在工作室中央的小凳子上,哼着歌踮脚用大画笔往空中“画画”。年糕坐在下方黑条上,身体和一条长尾巴都画在黑条前面,尾巴垂下来轻轻晃动。黑条边上放着一罐青色颜料。
2. 00:01.5,年糕的尾巴扫到颜料罐,罐子翻倒弹出黑条,画在黑条前面,因透视被夸张放大,一道青色颜料呈弧线甩向镜头;桃桃倒吸一口气,头发炸起。
3. 00:03,桃桃飞身扑出。她的一只手臂和肩膀越过下方黑条,画在黑条前面,因透视被夸张放大,手在半空接住颜料罐;身体其余部分仍在黑条后面的工作室里,贝雷帽被带得歪掉。
4. 00:04.5,“啪”的一声,颜料罐里剩下的青色颜料糊在镜头玻璃上:画面左侧三分之一被一大片青色颜料覆盖,颜料向下流出几道滴痕。黑条保持不变,不被颜料覆盖。桃桃在颜料后面,表情呆住。
5. 00:06,桃桃用袖子在青色颜料上擦出一个圆形的透明圆洞,脸从洞里探出来,鼻尖沾了青色颜料,她眯着眼坏笑。她的脸和肩膀越过下方黑条,画在黑条前面,体型很大。
6. 00:08,桃桃坐在下方黑条上,两条腿垂在黑条前面轻轻晃,怀里抱着年糕;她举起画笔,在剩下的颜料上画了一个爱心,对镜头眨一只眼,竖起大拇指。
【关键】凡是越过下方黑条的物体(尾巴、罐子、手臂、脸、腿)都必须画在黑条前面、盖住黑条,形成“探出画面”的错觉;其余部分仍留在画面区域里。黑条本身始终不变。不要任何文字、水印、碎片、速度线。














