Should be fine. I have GTX 3060, 12GB VRAM and I use the full fp16 on both checkpoint and CLIP for all models.

There's a fp8 checkpoint though at half the size, if you want that. Link is in the description.

I find that the results gets better and prompt adherence a lot better with triple clip. But you don't have to use it, you can use dual clip if you want.

CLIPAttentionMultiply Input Parameters:

clip

This parameter represents the CLIP model that you want to modify. The CLIP model is a powerful tool that combines text and image embeddings to understand and generate content based on both modalities. By providing the CLIP model as input, you enable the node to apply the specified adjustments to its attention layers.

q

This parameter controls the scaling factor for the query projection weights and biases in the attention mechanism. Adjusting this value can change how the model interprets the importance of different parts of the input data. The value ranges from 0.0 to 10.0, with a default of 1.0, allowing for fine-tuning to achieve the desired effect.

k

This parameter controls the scaling factor for the key projection weights and biases in the attention mechanism. Modifying this value can influence how the model matches the query with the key, affecting the attention distribution. The value ranges from 0.0 to 10.0, with a default of 1.0, providing flexibility in adjusting the attention behavior.

v

This parameter controls the scaling factor for the value projection weights and biases in the attention mechanism. Changing this value can alter how the model processes the values associated with the keys, impacting the final attention output. The value ranges from 0.0 to 10.0, with a default of 1.0, allowing for precise adjustments.

out

This parameter controls the scaling factor for the output projection weights and biases in the attention mechanism. Adjusting this value can affect the final output of the attention layer, influencing the overall performance of the model. The value ranges from 0.0 to 10.0, with a default of 1.0, enabling detailed customization.