Diffuse Map Surfaces版本v5.0-Pinnacle (ID: 1184885)

Creates detailed surface textures without shadows or hotspots.

For Version 5, the recommended "Weight" /"Strength" /"LoRA Weight" (depending on the tool you are using) is between "100%" and "150%" when entered as a percentage or "1.0" - "1.5" when entered as a decimal. Some prompts did a lot better at 150%, while others did better at 100%. If you are unhappy with the results, I recommend trying a different weight. Weights closer to 1.0 tended to produce more dynamic output, but sometimes added undesirable elements; weights closer to 1.5 produced more true diffuse map output but are sometimes too plain.

The recommended sampler is: DDIM

The Minimal Weight /Strength is: 1.0

The recommended Weight /Strength is: 1.0 - 1.5

The recommended Guidance Scale is: 3.5

The recommended number of steps are: 50-150

For best results, begin your prompt with "A diffuse map of a "

Example Prompts:

A diffuse map of a wooden surface with a natural wood grain texture.

A diffuse map of a brick wall.

Please be advised that the trigger word used in versions below version 5 was “DiffuseMapSurfaces, ” However, the trigger word was dropped for version 5.

Version 5 actually has no official trigger word but if you review the labels in the dataset, most of them begin with “A diffuse map of a” (a few omit the “a” at the end, so disregard it if it does not align grammatically with your prompt).

Comparison of output with and without the LORA using the same settings, seed, and prompt:

Dataset:

The dataset used is available for download from my huggingface account at:

https://huggingface.co/datasets/alastandy/Diffuse_Map_Surfaces

@misc {andrew_smith_2024,

	author       = { {Andrew Smith} },

	title        = { Diffuse_Map_Surfaces (Revision 9217a1f) },

	year         = 2024,

	url          = { https://huggingface.co/datasets/alastandy/Diffuse_Map_Surfaces },

	doi          = { 10.57967/hf/3835 },

	publisher    = { Hugging Face }

}

描述:

Version v5.0-Pinnacle:

Using 1.0 / (100%) weight/strength is once again working and recommended. In fact, in testing, using 1.0- 1.5 / ( 100% - 150%) produced some of the best version 5 results. Some prompts do better at 1.0 and some do better at 1.5.

Version 5.0-Pinnacle has a large and diverse set of training examples with labels. The dataset is 1.07 GB in size with hundreds of high-resolution images, and it took almost 24 hours to train 20 epochs on the civitai onsite trainer.

Start Time:

12/19/2024 03:14:17 a.m.

End Time:

12/20/2024 02:03:15 a.m.

I was honestly not sure if it would make it given the 24-hour time cutoff for failing a job. (Training processing for more than 24 hours is automatically failed by Civitai.)

训练词语: A diffuse map of