Huggingface: https://huggingface.co/sheldonxxxx/MechaDream

Support me on Ko-fi to help me buy more GPU time: https://ko-fi.com/techadventurex

MechaDream

A Stable Diffusion model for Mecha

All in one model (ckpt+lora+vae in one safetensors): https://civitai.com/models/47135

Recommended Setting

Checkpoint: Counterfeit-V2.5

VAE: mecha_v2_e3-pruned.ckpt (https://huggingface.co/sheldonxxxx/MechaDream/resolve/main/mecha_v2_e3-pruned.ckpt)

Sampler: UniPC (Personal preference)

CFG Scale: 15

Negative Embeddings: EasyNegative, bad-hands-5

Lora Weight: 0.8

Hires.fix with scale >= 2 is highly recommend

My workflow

Step 1: Batch generate images

Step 2: Select the best image and inpaint any flaws

Step 3: Upscale using MultiDiffusion

? Some comparison images between original image and after the above workflow ?

https://imgsli.com/MTcyNDM2

https://imgsli.com/MTcyNDM3

https://imgsli.com/MTcyNDM5

Roadmap

MechaDream v1 will be the base model for generic mecha image generation

New Lora models train on different anime Blu-ray will add flavors into my MechaDream

List of WIP anime:

1. Gurren Lagann (Dataset is ready, will start training when I'm less busy)

2. Code Geass

3. Macross Frontier

4. Evangelion

5. Genesis of Aquarion (Maybe)

6. Valvrave

7. Eureka Seven (Maybe)

8. Gundam

9. Armored Core

描述:

Base model: Counterfeit-V2.5

The Road to this Lora

Step 1: Fine-tuned on a dataset consisting of 130,000+ images of mecha/robots from various sources

Step 2: Mix the Lora models trained using different parameters

Step 3: Test different mixing ratios to find the best model

训练词语: