This model is different from the past. If you use it,please read the contents carefully:,

This model is a lightning big model pair (text emphasis/CFG Scale) comparison. The larger the inscription value,the higher the saturation. According to the official recommend and the CFG value "2" tested by myself,the expressive force is the most prominent and true,

The recommend range of the sampling step (Sampling steps) is between 10 and 30,and the best value "25" is obtained after the multi-vocabulary test ",

This model fixes and integrates VAE,keeping the VAE status as (None/None) in the running graph,

The amplification algorithm can use 4x-UltraSharp (high-score iteration step number "15"),

Sampling method recommend DPM 2S a,

This model is a multi-class large model can generate a variety of different styles,you can be bold innovation,

Must Read: V8.3 (Dark Photography Edition) belongs to the fine-tuning version, the training of the light shadow texture effect and the whole body hand crash problem to optimize and adjust with the big model fine-tuning synthesis, this model of the sampler to adjust the attention of the sampling ## Restart ##

If you like gray photography texture, you can use this version.

(This version can be used as a base film for training to understand the hands much better, and greatly reduce the crashing problem)

Support for image production (headshots, portraits, half-body, full-body)

Suggested sizes: 768, 1024, 1280, 1536 (2048*2048 images will have multi-body problems, if you want better texture, you can use the amplifier to realize large size images 4x-UltraSharp)

Iterative deployment: (25-30 steps)

Cue word guidance coefficient CFG scale: (7)

Recommended sampling method: Restart

Recommended Zoom Algorithm: 4x-UltraSharp (or) 4x_NMKD-Superscale-SP_178000_G

(Redraw amplitude recommended around 4) (High score iteration steps: 25 steps - 30 steps)

(Local map running - zoomable Recommended 1.5-2x)

About this version

Must Read: In V3, Retraining Using Cog Natural Language TAG Tagging Training

In the large model training set, extract some pictures for LOHR training and integrate them into the large model.

Output image support (headshot, portrait, half-body, full-body)

Iterative deployment: (25-30 steps)

Cue word guidance coefficient CFG scale: (7)

Recommended sampling method: DPM++ 2S a

In the V2 version, the character texture has been further optimized to enhance the detail and realism of the texture. At the same time, the light effect has also been optimized, so that the model can present a more realistic light and shadow effect under different lighting conditions.

This update makes changes to the Sampler. Please note the following recommended parameters

Iterative deployment: 25 -30

Recommended Sampling Method: DPM SDE Karras

Recommended magnification algorithm: 4x-UltraSharp (or) SwinIR_4x (redrawing amplitude is recommended to be around 0.3-0.5) (local operation chart magnification is recommended to be 1.5-2 times)

It is recommended to open the ADetailer run chart.

This model photography light and shadow sense has been fine-tuned, realistic visual effects

Iterative deployment: 25 -30 (30 details recommended skin is very good)

Recommended Sampling Method: DPM 2S a Karras

Recommended magnification algorithm: 4x-UltraSharp (or) SwinIR_4x (redrawing amplitude is recommended to be around 0.4-0.6) (local running map-magnification can directly open 2 times with good picture quality)

It is recommended to open AD running chart.