Gladiators - Fight or Love 2
Two gladiators wrestle each other at the beach. Will they fight or fall in love?
Video
Vertical (9:16)
Square (1:1)
Introduction
Having improved my knowledge and workflow working with Stable Diffusion in the past couple of months, I thought that I would revisit an earlier project, where I converted a Midjourney image into different variations using Stable Diffusion.
See: Gladiators - Fight or Love 1
Originally, I was going to work with the same image source image that I have made with Midjourney v5 (image below), but decided not to since it would look like a repost of earlier works.
The other option I considered was to remaster that image inside v5.2, but the results from the remaster were disappointing — with its increasingly strict content filtering, the muscular bodies of the gladiators are completely covered in body armor. While they can be visually pleasing, I really prefer the human forms.
So I went with a third option — I found an image generated on the same day but with a different seed number:
This image is not without issues — the left hand and the left foot of the gladiator on the right has some logic issues. But I went ahead with it anyway, thinking that with my workflow idea in mind, it shouldn’t create too huge of an issue. Ultimately, it did create issues in some of the renders, but not in a completely unusable way.
Workflow
Instead of using Softedge HED and OpenPose inside txt2img as I did in version 1, I used a workflow that I usually use for adding details and upscaling — with img2img and the Tile Resample control net. The only thing I didn’t do is apply the SD Ultimate Upscale script, which I would, if I wanted to do 4x to 8x upscale.
Using the 1024x1024 image created with Midjourney, I upscaled it for 1.5x to 1536x1536 during the process. I applied a low weight (0.5) in the control net and set it to Control Net-priority (aka “Control net is more important”), so that the new images will have some room to move the subjects around.
For the img2img setting, I applied 0.6 Denoise. I could set Denoising a lot higher and the results will still be usable. In fact, because of the presence of the Tile Resample Control Net, you would recognize the image even if it’s set to 1. But the images generated at 1.0 will be fairly different from each other. By this time, I already had an idea of what I wanted to do with those images, and it depends on smaller variance amongst the images. So I settled at 0.6 after looking at the early renders.
For the model, I used Virile Fusion v2, my favorite Stable Diffusion model at the moment for rendering men. I rendered these images at 50 steps with DPM++ 2M Karras. Yes, these settings are rather extreme, but that added a lot of details to the final images, which you will see in the variation comparisons below. I now have access to RTX 4090 thanks to vast.ai, and that allowed to me work more efficiently. I use a M2 Max with 96GB RAM, but since the current SD UIs don’t have full Apple Silicon support yet, my renders typically run at 3 IT/s, which is decent speed, but certainly no match to an RTX 4090.
Tools
As a summary, here are all the tools that I used to create the video.
- Midjourney v5 text prompt. 1024x1024.
- Stable Diffusion img2img, Denoising at 0.6, 1536x1536.
- Control Net: Tile Resample, Control Net priority, 0.5 weight.
- Virile Fusion v2 model by Scratchproof, 50 steps, DPM++ 2M Karras.
- Topaz Gigapixel HQ model at 2x, 3072x3072.
- Adobe Lightroom color correction.
- Adobe After Effects animation. A couple of indispensable plugins:
- Ableton Live for remixing the licensed music to match animation length.
Variations
With this workflow, you’ll see lots of details being added, with slight variations among the images. We’ll go through some examples below with a zoomed in view from different images.
Head Accessories (Gladiator on the Left)
Hair (Gladiator on the Right)
Red Garment
Blue Garment
Leg / Thigh Body Armor
Golden Cape
Waist
Accessories
All Variations
There is not a single perfect with the perfect variation. If I were to create a final image from this, then I would put all of these variations into Photoshop then manually mask our what I want to have in them, and maybe do a bit of inpainting. That would take quite a bit of time to do. I may work on it in the future, but for now, here are all 47 images generated with those the same parameters that I used to create my final video.
- 1Beach Fight 1v1
- 2Group Fight at the Beach (Study)