eth_10k

最强u卡-返现3%：

[eth_10k @gasfee_404]

@slazaruseth @paulism 牛逼，哥们。 还会用儿话音

[Fokki @0x_fokki]

A 20-year-old girl from Japan asked her boyfriend for $67. She never told him why.

By the next morning, her 18-second clip had reached a million views.

Her plan was simple: take the doll dress-up format already spreading everywhere and keep the same girl inside every outfit.

The trend wasn’t hers.

She borrowed the structure that was already working: close-up hands, a doll-sized figure, rapid outfit changes and a cozy room in the background.

The $67 paid for four subscriptions.

> Kimi K3 created the character file and wrote every shot prompt
> Seedream generated the stills
> Seedance animated only the images that passed the consistency check
> ElevenLabs voiced the character’s final line

That same video went live on TikTok, Reels, and Shorts in a single night.

Then she ran the system again: different outfits, a different room, the same recognizable face.

One evening of work.

One character she can keep building around.

[Rachel🥥 @Zesee]

最近很火的娃娃变装视频提示词开源

9:16 竖屏，超写实真人芭比换装短片。人物身份只参考上传的银灰发女孩图片，五官、脸型、眼距、鼻子、嘴唇、肤色与参考图保持一致。银灰色高马尾，轻薄空气刘海，两侧少量自然碎发。人物为真实成年女性，正常纤长模特比例，8至8.5头身，头部约占全身高度的八分之一，肩宽自然，四肢修长但符合真实人体结构，绝对不能大头小身。

人物始终位于画面正中央，全身完整入镜，双脚和头顶均有留白。平视固定机位，50mm镜头观感，无推拉、无摇移、无旋转、无焦段变化。背景固定为奶油白、雾霾蓝和少量香槟金组成的梦幻时尚工作室，柔和自然光，家具、墙面、地板和阴影位置全程不变。

画面前景出现一双真人尺寸巨型女性双手，双手皮肤真实，比例稳定，左右手结构正确。长杏仁形镜面银色美甲，镶嵌透明水晶、珍珠和精致立体蝴蝶结。巨手始终在人物前方活动，通过衣服、鞋子或配饰短暂遮挡人物，实现无缝换装。

最终服装、鞋子、项链、手链和手提包的款式、颜色、纹理，严格以对应的上传图片为准，禁止模型自行重新设计。 00:00–00:01
银灰发女孩穿基础内搭站在画面中央，身体自然放松，目视镜头。两只巨型美甲手分别从左下方和右下方进入，捏住连衣裙左右肩部或肩带，平稳地将裙子送到人物面前。

00:01–00:02
巨手同步向上抬起连衣裙，裙摆从人物膝盖位置上升到腰部和胸口。女孩低头看向衣服，身体位置、脸和发型保持不动。裙子逐渐遮挡基础内搭。

00:02–00:03
连衣裙继续向上移动，完整遮挡人物躯干，并在画面中央形成约0.3秒的自然遮挡。遮挡最完整的瞬间完成隐藏换装，不能用闪光、烟雾或溶解特效。

00:03–00:04
巨手向左右两侧缓慢分开并稍微下降。遮挡解除后，女孩已经完整穿好与母版图片完全一致的连衣裙，领口、腰线、裙摆和材质全部固定。

00:04–00:05
巨手轻轻整理裙子两侧肩部和腰线，布料产生细微自然摆动，随后双手从画面两侧退出。人物抬头看向镜头，保持真人芭比般克制、自然的表情。

00:05–00:06
两只巨手从画面底部左右两侧再次进入，各自捏住一只与鞋子参考图完全一致的鞋，鞋尖朝向人物双脚，缓慢送向脚边。

00:06–00:07
巨手将鞋子移动到人物双脚正前方，鞋子和手指短暂遮挡脚部。遮挡瞬间完成换鞋，双脚不能移动位置，不能穿模，不能出现多余鞋子。

00:07–00:08
巨手向两侧撤开，女孩脚上已经穿好成对的新鞋。鞋型、颜色、鞋跟和装饰左右一致。人物轻轻调整一次重心，裙摆有非常轻微的自然晃动。

00:08–00:09
硬切到固定的腰部以上中近景，不做连续镜头推近。两只巨手从画面上方左右两侧进入，拉直项链两端，将项链移动到女孩颈部前方。女孩微微抬起下巴。

00:09–00:10
巨手把项链绕到颈后，手指短暂遮挡颈部两侧。完成遮挡后，项链自然佩戴在颈部中央，吊坠位置端正，款式与项链母版一致。

00:10–00:11
巨手从上方退出，项链保持稳定，不闪烁、不变形。硬切回原来的固定全身构图，人物位置和背景透视必须与换装前一致。

00:11–00:12
女孩缓慢抬起左侧前臂，手掌向上。右上方进入一只巨型美甲手，拇指和食指捏着与参考图一致的手链，向女孩手腕靠近。

00:12–00:13
巨手将手链从女孩指尖方向轻轻滑向手腕，手指短暂遮挡手腕。遮挡解除后，手链已经正确戴好，手链大小贴合，不能悬浮或嵌入皮肤。

00:13–00:14
巨手轻轻调整手链位置后退出。女孩放松左手，同时略微抬起右侧前臂，为手提包动作留出空间。

00:14–00:15
一只巨手从右下方进入，捏住手提包的两个提手，将与母版完全一致的手提包抬到人物右侧前臂旁。包体保持真实重量感和稳定结构。

00:15–00:16
巨手将包带挂到女孩抬起的右前臂上，短暂遮挡手腕后松开。手提包自然下垂并轻轻摆动一次，不能穿过手臂或裙子。

00:16–00:17
巨手从右上方重新进入，拇指和食指像操纵真人娃娃一样，轻轻捏住女孩抬起的右手。女孩保持放松，不挣扎，手提包随手臂一起被抬高。

00:17–00:18
巨手将女孩右手略微向上提起。女孩像精致真人娃娃一样向前轻轻鞠躬约15至20度，头部、肩膀和上半身自然同步，银灰马尾和手提包轻微摆动。

00:18–00:19
女孩缓慢恢复直立，巨手松开并从右上方退出。人物面对镜头展示完整造型，双脚并拢，表情自然，保持最终造型约1秒，所有服装和配饰稳定不变。

00:19–00:22
不需要AI生成。这部分原片只是抖音搜索尾卡。剪辑时可以直接结束，或单独制作静态结束页。不要黑发、不要棕发、不要改变人物身份、不要换脸、不要欧美化五官、不要大头娃娃、不要Q版、不要塑料玩偶皮肤、不要缩短腿部、不要过细腰部、不要身体变形。

[John @johnAGI168]

剁腿变装👚

Seedance 2.5 prompt 👇

Duration: 10秒
Aspect ratio: 9:16
Overall style: 夏季时装卡点变装，真实摄影，一镜到底（One Take），动作匹配剪辑（Action Match Cut），4K高帧率

SHOT 1（00:00-00:02.20）
Subject: 女主@图片1。初始穿白色修身短衬衫、深灰高腰喇叭裤、黑色一字带细跟凉鞋。
Action:
女主从右侧走入，身体侧向左方，脸转向镜头，走到中央停下。
右腿承重；左膝轻屈，左脚向后勾起，随后快速跺回原位。双手屈于腰侧，手腕随音乐轻摆。
Environment: 白墙、黑色踢脚线、浅灰瓷砖地面，影棚柔光。
Camera: 固定全身镜头，双脚完整入画。
Constraints: 身体保持侧姿，脸和视线始终朝向镜头。

SHOT 2（00:02.20-00:08.20）
Action:
女主固定站在中央，反复使用同一动作：右腿承重→左膝轻屈→左脚向后勾起→左脚快速跺回原点。每次左鞋跟触地，服装瞬间硬切；换装前后脸、头发、侧身角度、手势完全衔接。

00:02.20：白色细肩带短上衣、白色侧抽绳短裤。
00:03.10：黑色半透短款上衣、金色亮片短裤、黑色丝袜。
00:03.90：灰色细肩带短上衣、灰色高腰紧身裤。
00:04.60：黑色高领短上衣、黑色短裤、红边黑色长筒丝袜。
00:05.25：紫色挂颈短上衣、黑色短裤、紫色丝袜、紫色腰带。
00:05.85：酒红露肩短上衣、黑色皮质短裤、黑色丝袜。
00:06.40：银灰单肩短上衣、黑色高腰短裙。
00:06.90：深蓝挂颈修身短裙、黑色丝袜。
00:07.35：象牙白缎面短上衣、灰色修身短裤、深灰丝袜。
00:07.80：白色修身短衬衫、黑色高腰皮质短裙。

Style: 后半段换装逐渐加速；每套都是贴身、利落的夏季造型。黑色细跟凉鞋全程不变。
Constraints: 每次只在左脚落地帧换装；不使用闪光、烟雾或旋转遮挡。

SHOT 3（00:08.20-00:10.00）
Action:
女主停止跺脚，保持侧身，脸看镜头；右腿承重，左膝微屈。她轻摆双手，随后微笑定格。
Camera: 固定全身镜头，结尾轻微推进。
Audio: 快节奏卡点音乐；每次左鞋跟落地配清脆“嗒”声。无对白、无字幕。

---
Negative: 日常通勤装，宽松T恤，卫衣，牛仔长裤，工装裤，运动服，厚外套，长裙，脸看向侧面，背对镜头，右腿跺地，左右腿交替，双脚跳跃，高抬腿，向前踢腿，脚底漂浮，换装早于落地，服装渐变，闪光，烟雾，人物瞬移，姿势断裂，额外腿脚，脸型变化，发型变化，鞋子变化，镜头晃动，背景变化，字幕，水印，标志。

[eth_10k @gasfee_404]

感觉显卡这波涨价又得持续个1-2年

[Flying puppy25 @df_reno]

前女友非要试试这个视角的，怎么样？#MinMaxH3 提示词：👇

[くろ@なんか刺さるAI美女 @ai6vyy]

こんなくだらん動画作れるのも、MiniMAXだとほぼ無料だから
seedanceも使い始めたけど、こんなのにコスト掛けれないｗ

画像と動画のプロンプト置いときます👇

[rS_ローカルLLM×投資 @rS_alonewolf]

今更ながらMiniMax H3のReference to Video
顔の再現がいまいちだった原因が分かった。
結構単純なこと

▫️課題
参照画像を動画に当てはめて生成する時、
動き始めると顔が別人になることがあった。

▫️原因
ref_image_size を match にしていた。

＜match＞
出力動画サイズに見合う大きさまで縮小
速い
顔の細部は落ちやすい

＜max＞
参照画像をより高解像度のまま使う
遅い
顔の再現は強くなりやすい

▫️対処法
ここをmax に変えると
明らかに元画像の顔の再現が強くなった。

[Maxi Joe @ChainMaxi_Joe]

我本以为15一年的蓝标已经天下无敌了

谁知道还有25一年的p+😅😅😅

[Tibo @thsottiaux]

Here is how to enable a 1M-token context window in Codex for GPT-5.6 Sol.

Even though we have tuned the context limit in Codex to be set optimally when it comes to performance and cost, this is a common ask, so here it is documented.

A larger context window lets Codex retain more code, tool output, and conversation history before summarizing older material. You need a model that supports it. And GPT-5.6 Sol, for example, has a documented 1,050,000-token window.

Open ~/.codex/config.toml and add or update these settings at the top level, before any [section] headers:

```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```

The first setting selects the model. The second tells Codex to use a one-million-token context budget. The third starts automatic history compaction around 900,000 tokens, leaving some headroom. Restart Codex client and start a new session after saving.

To try the configuration for a single CLI session without changing your defaults:

```
codex -m gpt-5.6-sol \
-c model_context_window=1000000 \
-c model_auto_compact_token_limit=900000
```

Have fun, but also know that we tuned the default carefully!

[meropan @renataro9]

ほんとですか！めちゃめちゃ嬉しすぎます😭✨
私プロンプト載せてなかったですね😂
こんな感じので作りました✨
subject_definitions:
<Subject 1> is the woman in reference image 1: a mature anime woman in right-facing profile, long black-to-muted-brown hair, red eyes, multiple silver ear piercings, black gloves, fitted black clothing, a red jacket over her shoulders, and a lit cigarette. Preserve her exact face, profile, hairstyle, earrings, hands, clothing, cigarette, and composed hard-boiled expression.

<Subject 2> is the hooded figure in reference image 2: an androgynous young anime outlaw facing forward, messy dark hair, glowing yellow eyes with red upper highlights, a vivid red hood and scarf, black tactical clothing, straps, and buckles. Preserve the exact face, eyes, hood silhouette, costume, proportions, and emotionless stare.

<Subject 3> consists of the hooded woman and the tan cat in reference image 3. Preserve the woman's exact side profile, straight dark hair, violet eye, gray hood, black tactical clothing, red shoulder panels, and her eye line toward the cat. Preserve the cat's appearance, collar, scale, position, and eye line.

summary:
[reference generation] Generate one complete 15-second 16:9 anime hard-boiled motion-design title sequence with native synchronized stereo sound. Treat character cinematography, editing, compositing, masks, graphic transitions, and kinetic typography as one unified generation. The visual language resembles a professionally designed crime-thriller opening title: disciplined editorial timing, layered two-dimensional typography in three-dimensional screen space, silhouette mattes, tracked text, animated rules, negative-space composition, optical impact frames, and controlled graphic transformations. Use only black, deep charcoal, blood red, and dirty ivory. All visible typography must be deliberately typeset, stable, correctly spelled, and integrated into the visual action.

retention_analysis:
<Subject 1>: fully_preserved identity and wardrobe; partially_preserved original environment.
<Subject 2>: fully_preserved identity, hood, eyes, scarf, and tactical costume; weak_reference for background.
<Subject 3>: fully_preserved woman, cat, clothing, profile relationship, and eye line; partially_preserved flowers and background.
The shared anime rendering, hard shadows, black-red palette, and nocturnal atmosphere are attribute_transfer.
Never merge the three characters. Never exchange their hair, eyes, clothing, accessories, or facial features.

detailed_description:
[Shot 1] Pure black screen. At the first metallic click, an extremely thin blood-red horizontal rule draws itself from the extreme left edge to the extreme right edge. The rule stops exactly across the center. The line vertically expands into enormous condensed uppercase typography reading exactly "REDLINE". The letters are dirty ivory, extremely tall, tightly tracked, and cropped by the upper and lower frame edges. The word remains perfectly spelled and geometrically stable. The interior of each letter acts as an animated matte: fragments of all three reference environments flicker only inside the letterforms. On a heavy bass impact, the letters rapidly collapse back into the single red line.

[Shot 2] At 00:02.200, the red line becomes a moving horizontal mask that reveals <Subject 1> from left to right. Preserve her right-facing profile. Build visible dimensional separation: large cropped letters reading exactly "NO ALIBI" move slowly behind her head, while smaller red typographic coordinates move in front of the distant background but remain behind her hand and cigarette. The camera makes a restrained slow push toward her red eye. She raises the cigarette slightly; the ember pulses once. Smoke rises and is pulled into thin vector-like curves that trace the contours of the letters. The phrase remains correctly spelled and does not cover her face. At the end, the vertical stem of the letter "I" expands into a full-height red panel.

[Shot 3] At 00:05.100, the red panel wipes laterally and reveals <Subject 2> centered in frontal close-up. The panel breaks into a modular grid of rectangular red and black graphic cells. Several cells show offset close-ups of the hood, buckles, scarf, and eyes, while the central cell keeps the complete face stable. The grid moves with precise mechanical timing rather than random glitching. Huge black-on-red typography reading exactly "NO MERCY" appears behind the central portrait. First "NO", then "MERCY", each word snapping into position on separate percussion hits. The camera performs a fast controlled push-in toward the eyes. As the eyes flare, the letterforms stretch horizontally, pass around the subject as foreground and background layers, then compress into a narrow vertical red bar. No facial deformation, duplicate eyes, costume changes, or uncontrolled digital glitch.

[Shot 4] At 00:08.100, the vertical bar becomes the separator between <Subject 3> on the right and the tan cat on the left. Preserve their original profiles and mutual eye line. The layout resembles a severe editorial poster brought to life. The camera drifts laterally with subtle parallax. The cat turns its head only slightly toward the viewer. The woman remains still except for one blink and a small movement of her hood fabric. Exact typography "03:17 A.M." assembles vertically along the separator bar, one character at a time, with clean mechanical alignment. Thin ivory registration marks and short red rules extend outward from the time display. Every character stays sharp and readable. A circular outline expands from the cat's eye, crosses the screen as a graphic iris, and becomes the next transition.

[Shot 5] At 00:11.100, the circular graphic creates three rapid typographic portrait compositions in succession. First: <Subject 1>'s eye and cigarette ember inside the hollow counter of a huge letter "R". Second: <Subject 2>'s glowing eyes framed inside two horizontal cuts through the letter "E". Third: <Subject 3> and the cat divided by the diagonal stroke of the letter "N". Each composition is identity-stable and lasts long enough to read as deliberate graphic design. The red line threads continuously through all three cuts.

[Shot 6] At 00:13.200, all portraits, red rules, grid cells, registration marks, and scattered letter fragments accelerate toward the center and assemble with precise magnetic motion into the final word "REDLINE". The final title is massive condensed dirty-ivory uppercase typography on pure black, with one thin red line passing horizontally through the exact optical center of the letters. Beneath it, smaller widely tracked uppercase typography reads exactly "TRUST NO ONE". Both lines are correctly spelled, flat to the screen, perfectly stable, and sharply legible. No additional words appear. Hold the completed title without movement from 00:14.200 until exactly 00:15.000.

overall_soundscape:
Native stereo sound designed as part of the edit. Begin with a low electrical room tone and a sharp left-to-right stereo drawing sound synchronized to the red line. Typography transformations use dry metal type impacts, paper-cut swipes, mechanical shutters, projector punches, and short sub-bass transients. Subject 1 includes cigarette ember crackle, controlled breathing, glove movement, and a faint neon transformer buzz. Subject 2 includes fabric tension, buckle movement, a low electrical eye flare, and tightly synchronized grid clicks. Subject 3 includes quiet night air, soft hood fabric, one restrained cat breath, and precise clock-like ticks for "03:17 A.M." The final assembly produces rapidly converging stereo clicks followed by one deep centered impact. Leave the final title hold almost silent except for a fading electrical hum.

non_diegetic_music:
A tightly edited industrial noir-jazz score at 96 BPM. Use muted upright bass, dry brushed snare, isolated prepared-piano strikes, low analog synthesizer, and restrained tape noise. Every major typographic snap lands on a defined musical transient. Increase rhythmic density from 00:11.100 through 00:14.200, then remove the drums instantly for the final title hold, leaving only a low sustained bass tone.

[R@aiaicreate @aiaicreate]

生成モデルMiniMax H3用の軽量化高速生成LoRAが公開されコミュニティで話題。
サイズが5分の1に削減されVRAMを節約できる点が好評。比較検証でもオリジナル版と画質に差異がないとされ、軽量化を歓迎する声。
#MiniMaxH3 #ComfyUI

URLはリプ⬇️

[brucexu.eth @brucexu_eth]

我终于知道为什么我的微信这么慢了。。。大家用 Clash 的时候可以检查一下。主要是 WeChat 的请求，被走了漏网之鱼，然后走了远程服务器，一来一回慢的一批。然后用 Process name 来加上 rule 直接 direct 就好了。

[Elon Musk @elonmusk]

@cognition Grok 4.7 will exceed all current models.

That said, Anthropic is a great company and will probably release improved models soon.

However, the SpaceX training corpus is so awesome &amp; unique that I would be shocked if any model is better at real-world engineering than 4.7.

[eth_10k @gasfee_404]

@thsottiaux codex在一个聊天窗口 长时间对话 等待时间会越来越久❗️

[eth_10k @gasfee_404]

@FiniYang 1088x1920分辨率 24fps 时长5s 20步，5070+32g，花费了6600秒生成的。

[T8star @T8star_Aix]

王炸！IndexTTS 2.5开源！免费整合包，8G显卡玩！最强语音TTS语音克隆，情感及音色的完美平衡版！MiniMax H3完美搭档，开源过... https://t.co/0swNbE71aJ 来自 @YouTube

#ai #comfyui #人工智能 #T8star #index25

[Future Vibes AI - Educator @FutureVibesAi]

Prompt -
First-person POV, ultra-photorealistic macro cinematic style. A real miniature human, only the size of an adult's palm, embarks on an adventure in a real residential backyard garden. He is not a toy, miniature model, CGI character, cartoon, or figurine. He has realistic skin texture, natural hair, detailed facial features, woven fabric clothing, authentic body proportions, and lifelike human movement.

The entire sequence is captured in one continuous uninterrupted shot with no cuts or transitions. Handheld macro cinematography with realistic shallow depth of field, natural focus breathing, subtle handheld camera shake, and immersive cinematic realism.

The tiny adventurer runs at full speed through towering blades of grass that feel like a giant forest. He weaves between damp soil, oversized pebbles, and sparkling dew drops, pushing aside enormous grass leaves as he moves forward.

He leaps onto a giant clover leaf, grabs the flexible stem, and uses its momentum to swing gracefully across a muddy patch before landing precisely on a gigantic ripe strawberry. He quickly climbs toward the top of the fruit using its textured surface for grip.

Suddenly, an enormous dew drop falls from a leaf above, crashing into the ground and sending a dramatic splash into the air. The impact knocks him off balance. He instantly grabs a bending blade of grass, which flexes naturally under his weight before springing back with realistic elasticity, launching him toward a massive blooming flower.

He lands safely on one of the flower's soft petals. At that moment, a beautiful butterfly flies past, and the powerful airflow from its wings nearly blows him away. He struggles to keep his balance and tightly grabs the flower stem.

Finally, he sits peacefully in the center of the flower, overlooking the entire backyard from his miniature perspective. Warm golden sunlight filters through the trees, creating beautiful dappled light across the scene. Smiling with satisfaction, he picks up a ripe berry and takes a joyful bite while admiring the breathtaking miniature world around him.

Visual Style: True-to-scale miniature illusion, realistic gravity, believable weight, accurate momentum, inertia, follow-through animation, natural contact shadows, flexible plant movement, realistic leaf bending, authentic water refraction, physically accurate motion blur, shallow depth of field, macro cinematography, ultra-realistic textures, cinematic lighting, photorealistic 8K quality.

Negative Prompt: No subtitles, no dialogue, no music, no CGI appearance, no floating, no teleportation, no scale changes, no cartoon style, no toy-like appearance, no unrealistic physics, no video game look.

[梭哈.AI @SUOHA_AI]

Codex 这个用法一定要改，结果会好一个量级

90% 的人都把 Codex 当神仙了，一上来就直接 /goal： 帮我写一个 App，实际上它很难理解你的具体需求

分享一条OpenAI研究员 @reach_vb 都在用的技巧：先让 Codex 做 research，再根据结论设置一个合适的 /goal，而不是一上来就直接 /goal

例如：

我要开发一个 XXX。 暂时不要写代码。 先帮我做 requirements gathering 和 research：
这个问题的核心约束是什么
有哪些成熟方案 / 同类实现可以参考
主要技术风险和容易踩的坑
最小可交付版本应该包含什么、排除什么

把调研结果整理清楚后，再根据结论 set 一个合适的 /goal

你也可以跟 Codex 对话：
1️⃣ 把需求和边界问清楚
2️⃣ 把现有方案和风险摸透
3️⃣ 再锁定一个真正可执行的 goal

等 goal 定准了，再让它开始写代码，结果会明显稳很多

[Trade off @tradeoff665]

3$ 3个月的蓝v你敢信？

前几天刷推看到个代充网站，便宜得有点离谱。出于好奇拿小号试了下，做好被骗 3 刀的准备，结果付完款 8 分钟左右蓝标真到了

链接：https://t.co/cO6PH6Qkw0

我也不清楚为什么这么便宜，什么渠道咱也不知道。纯分享不背书，大家自行决断