He suggests using Fal.AI to train a custom LoRA ( fal.ai/models/fal-ai/flux-lora-fast-training ) with at least 10 images of the subject. Then use this model to generate images ( fal.ai/models/fal-ai/flux-lora ) and increase their resolution using an up-res tool. Finally, you can now move on to animating them.
He details how to train a LoRA on Kling using at least eleven videos of your character. Admittedly, this pipeline is a little more involved. He also suggests FreePik as another option.
My take: basically, if you can imagine it, you can now create it.
The problem with a lot of image generators is that they love selfies: front-facing portraits. But what if you want a profile? Ben has a two-step work-around:
“Generate a close-up photo of your subject’s ear and then use the editor to zoom out and create the rest of the image.”
He explains:
“The reason this works is because what Midjourney needed was a pattern interrupt. Take advantage of its usual way to generate images by finding the usual way to generate an image with a more unusual focus. It’s better to choose a focus that is already often viewed from the angle we want.
focus on a ponytail if we want to see the back of someone’s head
use a receding hairline to see someone from straight above
focus on the back pocket of a pair of jeans if you want the…
I wouldn’t recommend looking up someone’s nostril (I mean it’s an angle that works but I just wouldn’t recommend it.)
The point is we can generate any of these things using extremely simple prompts and get very unusual angles to be seeing a person from. And then starting from there once we have the angle well defined we can simply zoom out and make our chosen feature less prominent by changing our prompt to something else and so in the new image the angle we wanted is extremely well defined not by tons of keywords but by the part of the image we already generated.”
This works for Expressions as well. He explains:
“If we start with a photo of just a smile or just closed eyes or just a mischievous smirk, Midjourney will spend all of its effort to create a high quality closeup version of the exact expression we wanted that now, in just one more generation, we can apply to our character by simply zooming out.”
“1080p Professional Mode: Kling 1.5 now generates videos at 1080p resolution when using Professional Mode. While it costs more credits, the output quality is significantly better and sets a new standard for AI video generation.
Motion Brush: Kling has introduced Motion Brush, a long-awaited tool in the AI video generation space. Currently, it’s only supported in the 1.0 model but will be available in 1.5 soon. Stay tuned!
End Frames: End frames have been introduced in the 1.0 model and are coming soon to the 1.5 model, allowing for smoother transitions and more control over your videos.
Using Negative Prompts: Improve your outputs by adding negative prompts to filter out undesired elements. Copy and paste the following negative prompts into your settings:
Of particular note is the emotion it’s able to generate.
Plus, Tim signals that Kling is about to add a full-featured Video Editor. Stay tuned indeed!
My take: of course, some will lament these advances. Yes, tasks that workers once spent their lives performing are now accomplished immediately. Looking at you, Medieval scribe, hot metal typesetter, telephone exchange operator. More job transformation is sure to come. We are well into the Digital Age and its promise is bearing increasingly wondrous fruit.
CyberJungle, the Youtube channel of Hamburg-based Senior IT Product Manager Cihan Unur, recently posted a great video on consistent generated characters.
There are lots of great insights in this 20-minute video. Two outstanding takeaways:
First: a prompting guide for Flux.1. At 15:28 he reveals three prompting styles: list, natural language and hybrid.
Second: a guidance guide for Flux.1. At 17:18 he shows Photorealistic and Cinematic images with a wide scope of guidance values. He posits:
“The essence of guidance setting is a compromise or a balance between photo realism and prompt understanding.”
My take: to me, too often generated images look over-the-top and so ideal, they’re unrealistic. The key seems to be dialing the guidance down to two. Who knew? Now, you do.
The goal here is to consistently end up with the same real person in multiple generated video clips.
“In this tutorial we’ll learn how to use the Flux image generator to train a custom AI model specifically for your own face and generate AI photos of yourself. Then we’ll animate those photos with the Kling AI video generator, which in my opinion generates the best AI videos right now.”
In a nutshell, the process is:
Create an archive of at least ten photos of your star
Upload this to the Ostris flux-dev-lora-trainer model on Replicate
Train the LORA custom image model and use it to generate key frames
Upscale these images on Magnific, optionally
Generate six second clips in Kling AI with these images
My take: it seems week by week we’re getting closer to truly usable generated video that rivals (or even surpasses) Hollywood’s CGI/VFX. Imagine being able to train more than one LORA model into Flux for Kling. I have it on good authority that that is just around the corner.
My take: a lot of people will immediately claim this is heresy, and threatens the very foundations of cinema as we’ve come to know it over the last one hundred years. And they would be right. And yet, time marches on. I believe some variation of this is the future of ultra-low budget production. Very soon the quality will surpass the shoddy CGI that many multi-million dollar Hollywood productions have been foisting on us lately.
Love it or hate it, as of August 2024, AI Video still has a long way to go.
In this video, AI Samson lays out the current AI Video Pipeline. Although there are a few fledgling story-building tools in development, full-featured “story mode” is not yet available in AI video generators. The current pipeline is:
Create the first and last frames of your clips
Animate the clips between these frames
Create audio and lip-sync the clips
Upscale the clips
Create music and SFX
Edit everything together offline.
It seems new platforms emerge weekly but AI Samson makes these recommendations:
My take: You know, the current pipeline makes me think of an animation pipeline. It’s eerily similar to the Machinima pipeline I used to create films in the sandbox mode of the video game The Movies over ten years ago:
