Location hidden
•
•Created byProfile pictureBrian Lovett
Profile picture
Brian LovettProfile picture@pixeldojo·4h

Heygen Avatar is back on Pixel Dojo: turn one portrait into a talking video

Heygen Avatar came back to Pixel Dojo yesterday. Give it one portrait and either a script with a voice or your own audio, and it turns that portrait into a talking video.


Here's how it works:


  • Portrait: upload one clear photo of a face. A front-facing shot with even light is the easiest place to start.

  • Script or audio: type what you want it to say and pick a voice, or upload your own audio track instead.

  • Extras: captions and a background are optional.

  • Cost: 2 credits per second of video. It holds an estimate based on your script when you start, then settles to the real length when the video is done.


Here's a quick test to start with:


  1. Upload a portrait of yourself or a character you've made on Pixel Dojo.

  2. Paste a short script, around 10 seconds, so the first try costs about 20 credits.

  3. Try this one, then swap in your own words:


Hey everyone, quick update from the studio. We just finished a new batch of product shots, and I want your honest take. Tell me which one you'd post first.


If you'd rather use your own voice, record those lines on your phone and upload the audio in place of the script.


Try it here:


Everything else that changed this week:


Post your talking portraits in the replies. I'd love to see who you bring to life.


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·1d

Train your own LoRA on Krea 2 Image, MiniMax H3 Image or the new MOSS-Image

Yesterday we opened LoRA training for three more image models on Pixel Dojo: Krea 2 Image, MiniMax H3 Image and the new MOSS-Image. If you want the same face, character, product or style in every picture, this is how you get it.


Here's what training costs per 1,000 steps:


  • MOSS-Image: 15 credits

  • Krea 2 LoRA trainer: 20 credits

  • MiniMax H3 Image: 35 credits


Once a LoRA is trained, you can use it on MiniMax H3 Image and MOSS-Image. You can stack up to three at once and set how strong each one is. Krea 2 Image pictures made with a LoRA also come out cleaner now, and they run on Turbo.


MOSS-Image is a new image model in its own right too. It follows long, detailed prompts, and an image costs 0.5 credits at 1K or 1.5 credits at 2K.


Here's how I'd make a first LoRA:


  1. Pick 10 to 20 photos of one subject. Change the angle, the light and the background, but keep the subject the same.

  2. Write one caption with a trigger word you'd never type by accident, like "brvx jacket".

  3. Train it, then add your LoRA on MOSS-Image and try a long prompt like this one:


brvx jacket, worn by a hiker on a windy ridge at golden hour, low clouds rolling through the valley behind, 35mm photo, natural skin texture, soft rim light, shallow depth of field


Swap in your own trigger word and keep the rest.


Full details are in the changelog:


Photo and caption tips for training:


Post your first LoRA results in the replies. I'd love to see what you train.


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·2d

How to get exact text in your images with Qwen Image 2.1 (half a credit at 2K)

If you make posters, thumbnails or product shots with words on them, give Qwen Image 2.1 a try this weekend. It sets the words you write as clean type, and it renders at native 2K.


Pixel Dojo has two versions:


  • Qwen Image 2.1 Turbo: 8 steps, 0.5 credits per image. It's my everyday pick for drafts and volume.

  • Qwen Image 2.1: the full 40 step model, 1.5 credits. Use it for final art and the finest small print.


Here's what makes the text come out right:


  1. Name the piece first (a concert poster, a product hero, a thumbnail)

  2. Then the subject, the setting, the light and the finish

  3. Put every word you want in the image inside quotes, and say where it goes


Here's a prompt to copy into Turbo at 2:3:


Typographic concert poster, bold condensed letters "SALT AND STATIC" stacked across the top, a lone saxophonist silhouetted against a huge coral moon over a quiet harbor, screen-print grain, deep navy and coral palette, small text at the bottom reading "AUGUST 14 TO 16, PIER 9"


Swap in your own title and date, and keep the rest. You can also attach up to 10 reference images to edit or combine them, and stack up to 3 of your own LoRAs.


More example prompts are in the , or jump straight in on the . Post what you make here. I want to see your posters.


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·3d

Clipper is live: drop in a long video, get vertical clips ready to post

If you've got podcasts, streams or interviews sitting on a drive, this one's for you. Clipper is now in Pixel Dojo, under Quick starts in Studios.


Here's how it works:


  1. Open

  2. Upload a podcast, stream or interview

  3. Clipper finds the best moments and cuts each one into a vertical clip


Every clip stays framed on whoever is talking, and when two people are on screen, both are shown. Captions get added word by word, so the clips are ready to post.


It costs 3 credits per minute of video, and you're only charged once your clips are ready.


One thing to try first: pick a 10 to 20 minute episode where you and a guest go back and forth. That's where the two-person framing and word-by-word captions really shine.


Full details are in the . Post your best clip here, I'd love to see what you make.


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·4d

You can now add Pixel Dojo to ChatGPT and get images and videos right in the chat

You can now add Pixel Dojo to ChatGPT. Add it from ChatGPT's plugin directory, sign in once, and then just ask for what you want. There's nothing to install and no API key.


Setup is three steps:


  1. Open pixeldojo.ai/chatgpt and click Add to ChatGPT, then Connect

  2. Sign in to your Pixel Dojo account and click Approve

  3. Type what you want, and the image or video shows up in the chat


You can ask for images, videos, music and product ads. You can also turn a photo into a video, edit an image with words, upscale, add captions, and use the characters you've already saved in Pixel Dojo.


It uses the credits in your Pixel Dojo account, same as the website, and if a job fails you get those credits back. ChatGPT picks a model for you, or you can name one, like Seedream 5.5, Nano Banana 2.1, Seedance 2.5 or Kling 4.0.


Here's a good one to try first:


Make a cinematic 5 second video of a fox running through snow at sunset.


Then drop in one of your own photos and ask: "Bring this photo to life. She turns to the camera and smiles."


Two things to know. Links in the chat stay live for one hour, so ask ChatGPT to save the ones you want to keep (they go to My Library). And films aren't in ChatGPT yet, so for those, use Film Studio or Sensei on the site.


The video with this post is our 40 second demo of how it works.


Add it here:


What's the first thing you'll ask it for? Let me know in the replies.


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·4d

Nano Banana 2.1 vs Pro and 4 other models: 27 seconds and 1.5 credits against 42 and 4.5

Here's a quick review of Nano Banana 2.1, which is live on Pixel Dojo.


We ran one prompt on six models: Nano Banana 2.1, Nano Banana Pro, Seedream 5.5, GPT Image 2.6, Krea 3 and FLUX 3 Image. The prompt asked for two described people and exactly three plates. All six got the people right. Four set three plates, and Nano Banana Pro and FLUX 3 Image set two.


2.1 took 27 seconds and 1.5 credits. Pro took 42 seconds and 4.5 credits, and its picture had the warmer light. This was one try each, so results will vary.


It makes images at 1K, 2K or 4K and takes up to 14 reference images. For a consistent person, attach one photo and name what to keep. For edits, count your changes, then list what stays.


Price: 1 credit at 1K, 1.5 at 2K and 3 at 4K.


Google itself says it can struggle with small faces and fine details.


The verdict in the video: pick Nano Banana 2.1 for layouts, consistent characters and counted edits. Reach for Pro when one hero image matters more than the price.


Try it:


The prompting guide:


The full video (2:47), with the prompts in the description:


Which one would you reach for?


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·4d

Kling Image 4: native 8K, a matching series and a product page from one photo

A quick review of Kling Image 4, which is live on Pixel Dojo.


It draws at native 2K, 4K or 8K, takes up to 10 reference pictures, and has four modes: image, series, product page and website. Every new image in this video was made with Kling Image 4 on Pixel Dojo, one try each.


Prompt it in five steps: the subject, the setting, the light, any words in quotes, then the camera. For a series, number the frames and then say what stays the same. Our four-frame backpack ad kept the same man, jacket and backpack, and cost 7 credits at 2K. For a product page, attach one product photo and list the sections you want.


Against Kling Image V3 on the same prompt, version 4 stayed closer to the brief and finished in 40 to 47 seconds, versus 109 to 152. Against Seedream 5.5 it was close. Kling held the overhead angle, while Seedream drew better icons and costs less.


We timed our 8K portrait at 2 minutes 12 seconds.


Pricing is 1.75 credits at 2K, 2.5 at 4K and 4 at 8K. Product pages start at 6.5.


Where it slips: it costs more, 8K is slow, our product page skipped two sections we asked for, and pages stop at 2K.


The verdict from the video: pick Kling Image 4 for big prints, matching sets and product pages. For everyday single images, Seedream 5.5 costs less.


Try it:


The prompting guide:


The full review:


Which of those modes would you try?


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·5d
Pinned post

Kling Image 4 is live on Pixel Dojo: 8K images, clear text, and 2 to 10 image series

Kling Image 4 went live on Pixel Dojo today, the same day Kling announced it.


Here's what it does:


  • Sharp images at 2K, 4K or 8K, with clear text in seven languages

  • Series mode: 2 to 10 images that keep the same people, place and style. Handy for storyboards, lookbooks and comics

  • Product page mode: add one photo of your product and get a full product page back as one tall image. Website mode makes a website design the same way

  • Up to 10 reference pictures to guide it

  • It's also in Canvas and on the API


Pricing is 1.75 credits at 2K, 2.5 at 4K and 4 at 8K. Product pages start at 6.5 credits, and the button shows the total before you start.


The picture with this post is a Kling Image 4 infographic we made on Pixel Dojo today. Text is the part I most wanted to check, so here's a text-heavy prompt to try:


Infographic poster titled "HOW BREAD IS MADE" with four illustrated steps labeled "MIX", "RISE", "SHAPE" and "BAKE", each with a one-line caption and a small icon, warm cream background, clean bold sans serif type, the text must be exact


After that, try Series with 4 images of the same character in four places and see how well it holds the face.


What's new this week:


Find it with the other image models:


Post what you make in the replies. I'd love to see your series runs.


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·6d

How to prompt Muse Image: one photo prompt, plus mixing up to 10 pictures

Today's how-to is on Muse Image, Meta's image model we added on Friday. It makes 2K images in 8 shapes for half a credit each, and it can also edit a photo or mix up to 10 of your pictures into one scene.


The picture with this post is one of the examples from our Muse Image guide. Here's the prompt we used (4:3):


Overhead shot of a steaming bowl of tonkotsu ramen on a dark wooden counter, glossy soft egg cut in half, charred pork slices, scallions and nori, chopsticks resting on the rim, a splash of chili oil on the side, warm side light, steam catching the light


The guide suggests building your prompt in this order:


  1. Subject: who or what, with concrete details

  2. Setting: somewhere real, with one or two mood details

  3. Light: a direction, a color and a time of day

  4. Camera or craft: a lens or film look, or the medium for art

  5. Finish: what it's for, like an ad, a film still or a book cover


Write full sentences, not a pile of tags. Muse Image thinks through your prompt before it draws, so clear sentences give it more to work with.


To combine pictures, add them to the composer and refer to them as image 1, image 2, image 3. Say what should change, then what has to stay the same. For example:


The woman from image 1 holding the water bottle from image 2 on a sunny rooftop terrace. Keep her face and the bottle's label unchanged.


One thing that trips people up: with pictures attached, the result takes the shape of your first picture, so put the one with the framing you want first.


Try it:


The full guide, with six prompts you can copy:


Post what you make in the replies. I'd love to see what you come up with.


Brian

Profile picture
Brian LovettProfile picture@pixeldojo·Nov 15

Seedance 2.5 Flash vs Seedance 2.5: same six prompts, 95 seconds against 199

Quick review of Seedance 2.5 Flash, which is live on Pixel Dojo.


We ran the same six prompts on Flash and on Seedance 2.5, one take each: a spoken line, action, a product shot, anime, a drone flight and a dancer. Flash finished faster on all six, 95 seconds on average against 199, for 20 credits instead of 30.


It uses the same prompt language as Seedance 2.5 and makes clips from 4 to 30 seconds, with sound and spoken lines. The tradeoff is that it stops at 720p, and it takes 10 reference images where the full model takes 30.


Pricing: Flash is 2 credits a second at 480p and 4 at 720p. Seedance 2.5 is 6 at 720p and 16 at 1080p.


How we'd use them: draft on Flash, then render the keeper on Seedance 2.5 with the same prompt, at 1080p.


For the prompt, write one line: the shot and camera move, the subject and one action, the place, the light and lens, then "Sound:". For speech, put the line in braces and say who speaks, how, and in which language.


Try it:


The side by side on all six prompts:


The prompting guide:



What would you draft on Flash first?


Brian