Sign In

TD-UltraReal

372
3.5k
32
0
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
3,452
Uploaded
Jul 9, 2023
Base Model
SD 1.5
Training
Steps: 20,000
Epochs: 15
Usage Tips
Clip Skip: 1
Hash
AutoV2
435D49606F
0
0
0
0
0

TD-UltraReal

We're trying to hone in on the most dramatic photorealistic styling possible

A general purpose model focusing on ultra photorealistic styling, perfect for people and figures, but also on animals, environments, and inanimate objects like cars, ships, and even imaginative things like cyborg monsters etc. all with the goal of making it feel like it all happened IRL.

Available now on www.thinkdiffusion.com

Join and ask any questions/comments in our Discord Server

Notes:

  • This model has had extra training for hands. The model is able to produce 5 finger looking hands with ease by default. You don't have to prompt any special tokens. Keep in mind this is not always perfect and it varies with seed (in such cases, you may add bad_hands or similar trigger words to the negative prompt for additional correction).

  • Also, with proper prompting you are able to create images where you can set multiple different colors per item without color bleed (ex: background, hair, dress, eyes), as seen in the images below. This is another unique feature of this model.

  • Stable resolution up to 768x1024 or 1024x768 (if you encounter duplicates, change sampler to euler_a or unipc) or lower to 704x960

  • Suggested settings section (a clipskip of 1 is recommended, vae setting of vae-ft-mse-840000-ema-pruned.safetensors