AI TIPS #7: Your AI Can Spell. You Just Have to Ask Right.
Quick takeaway: if your AI image has wrong spelling, fix the prompt first.
I tested a cheap model [Z Image]. Same settings, two prompts.
Loose prompt: "Chrysanthuum" and fake words.
Template prompt: clean text.
It did not work every time. 2 of 5 tries were clean.
Strong models like GPT Image 2.5 made fewer mistakes even with a loose prompt.
Loose prompt (before):
make a cool poster for a jazz night that says Chrysanthemum Philharmonic Rendezvous, saturday at the old warehouse, with a saxophone player and neon lights
Template prompt (after):
Poster for a jazz night, a saxophone player on a dark stage with pink and blue neon lights.
Text: "CHRYSANTHEMUM" in bold sans-serif, white, top center, large.
Text: "PHILHARMONIC RENDEZVOUS" in bold sans-serif, pink, under the first line, medium.
Text: "SATURDAY" in bold sans-serif, white, bottom center, small.
Spelling: C-H-R-Y-S-A-N-T-H-E-M-U-M P-H-I-L-H-A-R-M-O-N-I-C R-E-N-D-E-Z-V-O-U-S S-A-T-U-R-D-A-Y
Show each text once. No other text, no logos, no price tags.
Tips to remember
- Keep text short. One headline is easier than a full paragraph
- Put the exact words in quotes
- Spell names and rare words letter by letter
- Say "no other text" or the model may add its own
- Small text needs Quality [Medium] or [High]
- One wrong letter? Do an edit: "change only the word X to Y, keep everything else the same". Do not start over
Opinion:
If the image is a start frame for a video, I leave the text out.
I add text in my editor at the end.
If you have a better way to get clean text please leave a comment.
3
1
263
[Grok imagine High Quality (SOTA)] is being retired by xAI November 2nd.
This means platforms like Venice AI will also lose access to it.
This is a shame because it is the best model to build an original character from with the best realism.
The replacement Grok 2.0 is extremely nerfed in quality and barely useable for text-to-image creations.
You have 5 days left to make use of it on Venice before it’s gone. 😭
2
495
Viral French Influencer Video?
Same swap technique with WAN 3.0
🤖 Made with AI
Quick Tutorial: Character Swaps in Video
I'm creating this in response to Higgsfield's intense marketing on their "Viral AI Influencer" campaign.
Here I'm referencing a video from [audreyfrog33] on TikTok, please follow them
Quick takeaway:
- You don't need Genjutsu
- You don't need Seedance 2.5
- You don't need Higgsfield
All you need to is an understanding of the tools
I created this with WAN 3.0 because it was blocked by Seedance 2.5
Required Items:
- Character Sheet
- Video Reference
Character Sheet Prompt
-----------------------------------------------
professional photo character sheet in flat gray background #7B7B7C, left side as an extreme close up of the subjects face, right side is front and back view from head to toe.
Subject is: [ Japanese woman in her early 20's, long straight purple fading hair, arm tattoos, boss suit sleeves rolled up, gloves ]
-----------------------------------------------
Tutorial:
- Find a video to use for character swap
- Video must be less than 15 seconds
- Load the video into WAN 3.0 as a reference
- Load your character sheet into WAN 3.0 as a reference
- Set duration for 15 seconds or exact duration of video
- Set Aspect Ratio to match (9:16)
- Set output to 480p (use higher if you wish)
- Add prompt and create
Video Prompt:
[START PROMPT]
---------------------------
GLOBAL STYLE
Photoreal 30-second clip, 16:9. Same grade, lighting, and lens as the source video. Adults 21+ only. No text, no logos, no character-sheet grid.
SCENE
Replace the person in the source video with the woman from the character sheet. Keep her action, timing, and path. Do not keep the original face, hair, body, or clothes.
REFERENCE ROLES
@Video1 = motion path, timing, camera, and location ONLY. Do not copy the original person's face, hair, body, wardrobe, or identity.
@Image1 = character-sheet woman, face + body + wardrobe ONLY. Do not copy the sheet layout, turnaround poses, labels, crop, or background.
CHARACTERS
Woman, 21+, identity locked to @Image1: same face, hair, skin, proportions, and outfit from the sheet in every frame. One person only. No extra limbs, no face morph, no blend with the original actor.
LOCATION
Same room, props, and background as @Video1. Do not import the sheet backdrop.
FIRST FRAME AND BLOCKING
Frame 1 already in motion, matching the source action. She is fully the sheet character from the first frame, not a fade or morph.
SHOTS
[0-30s] She performs the same action and blocking as the person in @Video1. Camera matches the source. Identity stays locked to @Image1. Last 3 seconds hold the finished pose.
OPTICS
Match the source lens and depth of field. No beauty filter.
CAMERA
Follow @Video1 camera path only. Do not restage.
PHYSICS
Weight, cloth, and hair match her real motion. No sliding feet, no rubber limbs.
LIGHTING
Match the source practicals and shadows. Her face takes the same light direction as the original person.
AUDIO
<same diegetic room and movement SFX as the source>
no music, no score, no subtitles, no on-screen text
CONSTRAINTS
Do not copy the original person's identity. Do not paste the character sheet as a picture-in-picture or grid. Do not change the action. Faces stay on-model. No text.
-------------------------------------------------
[END PROMPT]
Quick Tutorial: Character Swaps in Video
I'm creating this in response to Higgsfield's intense marketing on their "Viral AI Influencer" campaign.
Here I'm referencing a video from [audreyfrog33] on TikTok, please follow them
Quick takeaway:
- You don't need Genjutsu
- You don't need Seedance 2.5
- You don't need Higgsfield
All you need to is an understanding of the tools
I created this with WAN 3.0 because it was blocked by Seedance 2.5
Required Items:
- Character Sheet
- Video Reference
Character Sheet Prompt
-----------------------------------------------
professional photo character sheet in flat gray background #7B7B7C, left side as an extreme close up of the subjects face, right side is front and back view from head to toe.
Subject is: [ Japanese woman in her early 20's, long straight purple fading hair, arm tattoos, boss suit sleeves rolled up, gloves ]
-----------------------------------------------
Tutorial:
- Find a video to use for character swap
- Video must be less than 15 seconds
- Load the video into WAN 3.0 as a reference
- Load your character sheet into WAN 3.0 as a reference
- Set duration for 15 seconds or exact duration of video
- Set Aspect Ratio to match (9:16)
- Set output to 480p (use higher if you wish)
- Add prompt and create
Video Prompt:
[START PROMPT]
---------------------------
GLOBAL STYLE
Photoreal 30-second clip, 16:9. Same grade, lighting, and lens as the source video. Adults 21+ only. No text, no logos, no character-sheet grid.
SCENE
Replace the person in the source video with the woman from the character sheet. Keep her action, timing, and path. Do not keep the original face, hair, body, or clothes.
REFERENCE ROLES
@Video1 = motion path, timing, camera, and location ONLY. Do not copy the original person's face, hair, body, wardrobe, or identity.
@Image1 = character-sheet woman, face + body + wardrobe ONLY. Do not copy the sheet layout, turnaround poses, labels, crop, or background.
CHARACTERS
Woman, 21+, identity locked to @Image1: same face, hair, skin, proportions, and outfit from the sheet in every frame. One person only. No extra limbs, no face morph, no blend with the original actor.
LOCATION
Same room, props, and background as @Video1. Do not import the sheet backdrop.
FIRST FRAME AND BLOCKING
Frame 1 already in motion, matching the source action. She is fully the sheet character from the first frame, not a fade or morph.
SHOTS
[0-30s] She performs the same action and blocking as the person in @Video1. Camera matches the source. Identity stays locked to @Image1. Last 3 seconds hold the finished pose.
OPTICS
Match the source lens and depth of field. No beauty filter.
CAMERA
Follow @Video1 camera path only. Do not restage.
PHYSICS
Weight, cloth, and hair match her real motion. No sliding feet, no rubber limbs.
LIGHTING
Match the source practicals and shadows. Her face takes the same light direction as the original person.
AUDIO
<same diegetic room and movement SFX as the source>
no music, no score, no subtitles, no on-screen text
CONSTRAINTS
Do not copy the original person's identity. Do not paste the character sheet as a picture-in-picture or grid. Do not change the action. Faces stay on-model. No text.
-------------------------------------------------
[END PROMPT]
🤖 Made with AI
I wanted to see how WAN 3.0 handles TikTok style dance performances.
The first video I hand wrote the prompt, two videos that I took to TikTok to dub the original music.
The second video I used a skill I created for Grok to optimize it.
- Video Model: [WAN 3.0 Reference]
- Image Model: [Grok Imagine High Quality (SOTA)]
- Image Edit Model: [Seedream V5 Pro]
- WAN 3.0 Skill: grok.com/skill-link/38992646…
- Platform: Venice.ai
- Referral Link: venice.ai/chat?ref=4glZT-
Prompts in comments
Original Inspiration Post teapawt.org/LoveUolanda/status/209…
🤖 Made with AI
调试了9个版本终于稳定下来了。
这种网红视角的短视频在h3模式使用图生视频来完成 对提示词要求非常高。
还踩了个坑 .
H3 在高像素模式下,模型会优先追求人脸细节还原;俯拍 45‑55° 向下拍摄,人脸会产生透视形变(额头大、下巴收)。模型内部有隐性偏好:高清晰度生成时尽量避免人脸强烈透视畸变,主动把相机拉回眼平,保证人脸 “好看端正”,直接无视你的机位硬约束。
也在文章分享了修复方案。
提示词 teapawt.org/LoveUolanda/status/209…
芙宁娜|时装走秀 💙🎭✨
芙宁娜登上属于她的时尚舞台。
随着灯光与音乐展开,优雅、戏剧感与舞台气场交织在一起,属于水之国的时装秀正式开始。
从舞台中央走来,每一步都像一场属于芙宁娜的精彩演出。✨
今天的主角,当然是芙宁娜。
喜欢这场走秀的话,点赞❤️收藏⭐关注🔔支持一下!
推荐 X 标签:
#芙宁娜 #Furina #原神 #GenshinImpact #时装走秀 #FashionShow
#フリーナ #原神 #ファッションショー
#NguyenThan #GenshinImpactVN #NhanVatGame
#AIArt #AIVideo #提示词分析 #promptshare #MINIMAXH3 #ai教程
Image Prompt [Grok Imagine High Quality (SOTA)]
-----------------------------------------------------
A professional female Chinese genshin impact cosplay in white void studio, high angle camera. Model is petite skinny and tone. The specific genshin impact cosplay is Ganyu but a sexy version with black thigh high stockings
Image Edit Prompt (Character Sheet) [Seedream V5 Pro]
-----------------------------------------------------
professional female character sheet, framed from knee to head standing at ease in white void with same lighting, left side is front, right side is back
WAN 3.0 Video Prompt (480p, 30s)
-----------------------------------------------------
@Image1 and @Image2 are the same person Jill. @Image2 is the identity reference locking her identity. @Image1 is the scene, a high angle shot of a cosplayer performing a TikTok style dance recorded on an iPhone at a high angle mounted on a monopod in one continuous shot. Jill performs a TikTok style 130 bpm dance for 25 seconds.
WAN 3.0 Skill Optimized Prompt (480p, 30s)
-----------------------------------------------------
@image2 is Jill, the identity lock: same East Asian woman, same face, same dark plum lipstick, same light-blue contacts, same slim toned body and left-hip tattoo. Use the left half of @image2 for her face and front costume; use the right half only for the back of her hair, cape, and skirt. @image2 is a front-and-back reference sheet of one person — do not generate two people and do not copy the split layout into the video.
@image1 is the same Jill in the scene: white cyclorama studio, high-angle iPhone framing, costume and lighting as seen from above. Start from her @image1 pose — left hand on hip, right hand touching the left red-and-black horn, weight on her right foot, looking up at the lens.
Keep her costume locked for the full clip: navy hair in twin red-and-black horn buns with hanging gold bells, long purple-tipped ponytail tied with a purple ribbon, white fur collar with a gold bell and purple tassel, purple-and-gold embroidered keyhole crop top, black fur-trimmed cape with gold linework, long black gloves, short white-and-purple geometric wrap skirt with gold trim and tassels, purple thigh garters with gold diamond charms, black sheer stockings with purple geometric marks and lace tops, barefoot on the white floor. A small purple four-point emblem card stays on the floor near her feet, same as @image1.
Jill performs one continuous 25-second TikTok dance at 130 BPM, always in time with the kick.
0–3s: holds the @image1 starting pose for one beat, then drops her raised hand and hits the first downbeat with a sharp hip pop toward camera.
3–7s: side-to-side weight shifts and hip sways, shoulders isolating opposite the hips, one gloved hand pointing at the lens then snapping back to her hip; horn bells and skirt tassels swing on the beat.
7–11s: slow body roll down through the ribs and up again, ponytail and cape flaring; she glances up at the camera with a small smirk on the roll-up.
11–15s: compact grapevine / box step in place so she stays framed, arms waving in opposing figure-eights, mini-skirt flaring on each lateral step.
15–19s: half-turn so the camera briefly reads the back of the cape, purple ribbon, and skirt, two sharp hip pulses, then she whips back to face the lens with the ponytail following a beat later.
19–23s: faster foot taps and knee pops, both hands framing her face then sliding down the crop top to the skirt hem; bells jingle on every hit.
23–25s: she plants her feet, pops back into a high-angle end pose close to @image1 — one hand at a horn, the other on her hip — and holds the last two beats looking straight up into the phone.
Setting: bright even white cyclorama, no set dressing except the purple floor card. Single continuous take, no cuts.
Camera: iPhone mounted high on a monopod, high angle looking down, full body plus floor in frame, 9:16 vertical, slight ultrawide phone distortion at the edges. Monopod is mostly locked; only tiny natural phone micro-sway. No operator in frame, no gimbal glide, no orbit.
No dialogue. No on-screen text, no subtitles, no captions, no UI.
Audio: foreground 130 BPM punchy electronic-pop TikTok dance track with a hard kick on every beat; costume gold bells jingling in rhythm; soft stocking foot taps on seamless paper; light rustle of cape fur and skirt fabric. No voiceover, no extra speech.
1
3,246
WARNING ABOUT GROK AUTO TOP-UP
If you disabled it before you need to disable it again.
They made a change and enabled it by default again.
The new setting charges you if your balance of extra credit goes below a 3 dollar threshold. I was charged 5 dollars without any desire or consent on the setting.
This is predatory design
- The inability to remove payment method from top up unrelated to account sub.
- Enabling auto top up without user consent/warning
- Enabling auto top up again after previously disabling it
Visura retweeted
Rolling out over the next few days, you can decide if your image, video, and music creations made with Gemini will have visible watermarks. Invisible SynthID watermarks and C2PA metadata will always stay embedded in the background.
Simply go to your settings and toggle “Show watermark” on or off to set your preferences across all future images, videos, and music tracks.




















