▒░|:. Asheem Gujadhur .:|░▒ Original work. I post about my AI experiences. I create my own synthetic female selfies using my face. Goal: Realistic AI character.
Google Veo 3.1 nailed this oversized romantic Paris sunset video. It is now my phone's new lockscreen. #veo3#veo31
Process: Text to video, 1080p 8s.
Platform: @freepik (prompt published there)
Seedream 5 Pro marketing says that it can do pixel-level precision "AI Edit."
I used my Samsung S24U stylus to draw on problem area (left img). Uploaded to #Capcut App with prompt: "Follow scribbled instruction in image."
Repaired image (right). #Seedream changed the pose so maybe "Keep everything else the same" should have been added to the prompt.
Below is use of a more Asian/Middle-Eastern selfie that I created. It has slightly more expressive base face.
Same 4 facial expressions as before (grid-style, so quality is reduced): {stoic, smirk, happy, flirty}.
Testing 4 facial expressions in Seedream 5 Pro. My first time using a grid-style to display results. 2x2 grid {stoic, smirk, happy, flirty} prompt below.
Model: #Seedream 5 Pro #Seedream5
Platform: #Capcut App
Upload selfie + use img2img prompt:
"Front facing camera. Candid photography. Use face, make-up but not original hair from reference image to create the portrait of a slim woman with professionally styled long blonde hair. She is wearing a frilly French-maid dress. The background is a Victorian kitchen with natural light. A 2x2 grid to show 4 different facial emotions: (1) stoic face, (2) subtle smirk, (3) happy face, and (4) flirtatious face. Full-frame camera close-up shot with natural grain. Ensure facial consistency."
Good morning folks! Below, a jazz night scene involving a glamorous female singer on a piano & a pianist. Crikey, that #Seedream made my rough prompt look good.
Model: Seedream 5 Pro
Platform: @CapcutApp #Capcut
Use selfie + image-to-image prompt:
"Photorealistic cinematic portrait of a glamorous female jazz singer reclining on a glossy grand piano in an intimate late‑night jazz lounge. Use the face and makeup from reference image to create a tall, slim woman with long professionally styled blonde hair, wearing a red silk, figure‑hugging evening gown and red high‑heeled court shoes. She lies horizontally along the piano lid, legs crossed at the ankles, right elbow resting on the piano with fingers lightly touching her cheek, left arm extended with one finger pointing upward — relaxed, confident, and elegant. A Black pianist sits at the bench with his back to camera, hands on the keys, secondary in the frame, sharing a warm smile and gaze with the singer. Dim amber ambient light, soft smoke, shallow depth of field, gentle rim light on the singer, warm reflections on the piano lid, vintage Hollywood color grade, tasteful and evocative mood. Keep composition balanced so the singer is the clear focal point."
2-in-1 post:
- Left: @TopazLabs AI upscale of an img2img Nano Banana 1.0 image. More like overpriced dumbscaling.
- Right: Using that image as reference for Seedream 5 Pro Image Edit (i2i) (as offered by @capcutapp). Lots of positives.
#Capcut#Seedream#Seedream5
Tried #Seedream 5 Pro in @itsPolloAI App. It produces 2K images but the file size is a low 600KB. Still, you can see that #Seedream5 pro does more than apply a selfie when this close - realistic skin.
The Quote-Repost SD5pro gen is from yesterday in #Capcut App. Already, selfie application is different?
Same selfie & prompt used in image-to-image tests:
"Use facial identity, facial geometry and make-up but not original hair from [uploaded image] to create a fashionable blonde woman wearing a pale-blue denim tube top, red bead necklace and gold hoop earings. She is sat down at the outside table of a coffee shop, on a busy summer morning day in New York from year 2000. The view is entirely recorded from her phone. Low-angle shot aimed at her face and the far background includes the famous twin tower skycrapers and blue skies. One of her arms extends frame-right while the other lightly touches the rim of a half-empty coffee cup located in front of her. Foreground is her grogeous face with a subtle flirty smile, lips parted, long blonde side-parted hair with a healthy shine. Candid photography, natural grain and detailed shot."
My first "Seedream 5 Pro" test and it will be against "ChatGPT Image 2."
Nano Banana 2 Lite was binned in the reply.
Test: Image Edit (image-to-image) - Same selfie & prompt (see below)
Theme: Y2K
Platform: #Capcut App
#Seedream5 #ChatGPTImage2
Upload selfie + use prompt: "Use facial identity, facial geometry and make-up but not original hair from [uploaded image] to create a fashionable blonde woman wearing a pale blue denim tube top, red bead necklace, gold hoop earings and sat down at the outside table of a coffee shop, on a busy summer morning day in New York circa 2000. The view is entirely recorded from her iphone. Low-angle shot aimed at her face and the far background includes the famous twin tower skycrapers ans blue skies. One of her arms extends frame-right while the other lightly touches the rim of a half-empty coffee cup located in front of her. Foreground is her grogeous face with a subtle smile, lips parted, long blonde side-parted hair with a healthy shine. Candid photography and detailed."
I must admit that Seedream 5 Pro is pretty amazing. 😉 Image-Edit (i2i) prompt below.
Theme: Flirting at Christmas
Model: #Seedream in #Capcut App
#seedream5pro
Use a selfie + prompt: "Use facial identity, facial geometry but not original hair from the reference image. Her new hair is of a two-tone blonde colour, long, flowing and professionally styled. Create a high-fashion editorial Christmas scene featuring a woman flirting with a young and strong man dressed as Santa Claus. The setting is a lived-in indoor festive environment with a deep red carpeted floor, scattered Christmas ornaments, mirrored disco balls, and a partially visible champagne bottle on the floor. The woman is wearing a fitted deep-red, shimmering, metallic mini-dress, thin straps, and natural fabric tension around the waist and hips to accentuate her feminine beauty. Visible healthy skin texture and pores. She is sitting with slightly arched torso, in the lap of a tall and muscly man dressed in loose Santa outfit. One arm resting on Santa’s leg, the other holding a champagne glass. Lighting is dramatic, reflections from disco balls, flash-style editorial lighting with strong highlights and defined shadows, realistic reflections on velvet fabric and ornaments. Santa’s costume is rich red velvet with white fur trim, textured beard, round glasses catching light naturally. Shot on a full-frame DSLR, 35mm lens, shallow depth of field, crisp focus on faces, cinematic party, editorial style with natural grain."
#GrokImagine video 1.5 (10s + 6s extend) versus #Seedance2, using TechHalla's T2V prompt.
Theme: Christopher Columbus' ship.
The main difficulty was getting a generation where the crew was on the ship's deck rather than hanging precariously outside, about to fall at sea.
Who did it best?
- Nano Banana 2 (img in post)
- Seedream 5 Pro (img in quote-repost)
Read prompt before deciding.
#NanoBanana 2 (Gemini) v/s #Seedream 5 Pro.
Prompt:
"1990s Spain home video still frame, consumer camcorder (Hi8 / VHS-C), whitewashed Andalusian village street at siesta hour, handheld casual framing, imperfect focus, auto exposure shift, VHS tape noise, interlaced video look, low resolution, faded warm colors, overexposed sunlit walls and deep shade mixed lighting, dry dusty afternoon atmosphere, raw documentary footage look, unpolished realism, no cinematic grading, no stylization, no text, no watermark. Old woman in a dark housecoat sitting on a wooden chair beside her doorway in the shade, a dog lying on the ground at her feet, long shadows down the empty street, unaware of the camera, neutral mood."
@capcutapp > edit any part of an image with pixel-level precision
I used prompt to successfully repair an image. Is there another more convenient method? Thanks.
Using "Seedream 5 Pro" to AI edit and fix an error image.
Distorted leg is now straight and everything from lower dress, shadows were adjusted accordingly. Style of heel was retained, everything else constant.
#Seedream#Seedream5#Capcut
Just looking at new stuff on @itsPolloAI App. I was very impressed with their "Pollo Image 2.0" image-to-image model, set to "professional" mode for enhanced quality.
In terms of facial consistency & prompt understanding, Pollo Image 2 is up there. See results for yourself. 👍
Using Grok to correctly reverse engineer a source image into a prompt is hard. It does poorly even in expert mode.
Out of 20 daily #GrokImagine img allowance, only 1 was generated due to censorship. Face-to-image edits are heavily moderated even when fine. Nano Banana didn't b*tch one bit about the Grok prompt.