Return to Comfyui Workflows (also accessible via "Workshop" in the menu bar)

"I'VE GOT ALL THAT YOU NEED"

WORKFLOWS & PROJECT DETAILS

🎥 Description

A day in the life of a high-class escort. Set to Garage music. I went for a phone style vertical resolution to give it a more real-life feel and themed it on a 24 hour trip to Ibiza with a special kind of after-party. I blurred out the key bits and YT accepted it but made it 18. Rumble was fine with it. I haven't put the NSFW version up because Vimeo has run out of space on the free tier and I don't know where else would host NSFW (tell me if you know).

• Music: "I've Got All That You Need" by Mark DK Berry – Available on Bandcamp

• Date Published: 8th Feb 2025

🎬 About the Project

I applied a new Hunyuan workflow for this video, and it gave me better quality and was faster. This was due to switching out the previous Hunyuan model and lowering the steps while keeping the Fastvideo lora in. I also used Tea cache. I decided to set it in a different aspect ratio more suitable to phone devices this time for "realism".

It took about 4 days total, but I didn't have time to give it a color grading, which I plan to in future as video results get better. But I set myself a 5 day limit on the project and the end result was unlikely to benefit from colourisation.

The uncensored version of this song doesn't have a home yet (I ran out of free space on Vimeo with the last one). I am looking for another streaming service to offer it, but I am also not too concerned about NSFW content. I am fine sticking to SFW in most cases.

⚠️ Key Challenges

• Character Consistency

  The Lora worked well enough. I used a Lora that I got off civitai and thought she was just some random pawn actress and I would get away with it. I only later found out this was a celebrity. Oops.

• Legs, Hair, Body Shape

  Arms and legs tend to disintegrate, distort, and weird walking always happens. At the time it was just what the Hunyuan models did.

• Clothing Consistency

  It was easy enough to prompt for what little I needed.

• Psychological Toll

  Easy, and it was more fun because it was the first time I tried a Lora for character consistency and it worked okay. This was also a better and faster workflow than I had been using.

🔧 Workflows & Tools Used

• The below zip file contains x1 Comfyui json workflow:

  • hunyuan-fp8-with-fast-lora.json

  Right-Click and "Save link as" to download the ZIP containing Comfyui json workflows

⏱️ Time & Energy Investment

→ Total: 4 days

💻 Hardware

• GPU: RTX 3060 (12GB VRAM)

• RAM: 32GB

• OS: Windows 10

All of it was done on a regular home PC.

🧰 Software Stack

• ComfyUI (SDXL, Hunyuan t2v, inpainting models)

• Krita + ACLY plugin – Fast inpainting and upscaling

• Davinci Resolve 19 – Final cut

🎨 Loras Used & Trained

• Hunyuan FastLora - this added speed to the process of the hunyuan text-to-video model by reducing the steps required from +20 to 5.

• The female character was my first use of a character Lora. Don't recall the name, but it will be in the workflow. It was off Civitai and I didn't recognise the actress so I assumed it was AI created, but it wasn't. (I recommend avoiding using famous people as it could come back on you in the future and I wouldn't want my stuff stolen and used by others either).

📺 Resolution & Rendering Details

I dont recall what size I was making this, but it will be in the workflow. It was small, like 380 on the long side. I was going for speed over high quality just to get it done. You couldnt really get rid of the distorting limbs so there was no point going higher. They also didn't always follow the prompt, and I usually allowed three goes before picking the best of.

This was before Wan 2.1 model came out, and before image-to-video showed up on the scene. This was a text-to-video model. I also hadn't looked at interpolation or upscaling at this stage.

😵‍💫 Final Thoughts

It was fun, and I was really just mucking about with ideas to see how it went while waiting for a better model to be released. The Hunyuan text-to-video model had too many limitations before Wan 2.1 and image-to-video showed up.

back to top