From 8bc2b71ae4585b8f42a500374fdd8d91bb531350 Mon Sep 17 00:00:00 2001 From: bash-j Date: Sun, 26 May 2024 12:55:41 +0930 Subject: [PATCH] more tweaking to get better responses from llama 3 --- mikey_nodes.py | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/mikey_nodes.py b/mikey_nodes.py index 6a7b4ce..f512f06 100644 --- a/mikey_nodes.py +++ b/mikey_nodes.py @@ -4458,7 +4458,7 @@ class LMStudioPrompt: if mode == 'prompt': return { "messages": [ - { "role": "system", "content": "You are a helpful assistant." }, + { "role": "system", "content": "You are a helpful assistant who obeys the user's instructions." }, { "role": "user", "content": "I say something like 'blonde woman' and you respond with a single prompt which I can use as prompts for an AI txt2image model. Your response to blonde woman would be something like 'Blonde woman wearing a patterned orange, white, and blue sundress, smiling, on a sunny day with a blue sky, surrounded by buildings and palm trees'. Only respond with the prompt and nothing else."}, { "role": "assistant", "content": "Sure thing! Let's begin. What is your first prompt?" }, { "role": "user", "content": "futuristic man"}, @@ -4484,9 +4484,9 @@ class LMStudioPrompt: elif mode == 'pixelwave': return { "messages": [ - { "role": "system", "content": "You are a helpful assistant." }, - { "role": "user", "content": "I say something like 'photo of a blonde woman' and you respond with a single prompt which I can use as prompts for an AI txt2image model. Your response would be something like 'Blonde woman with a flower crown, digital color photograph, tags: flower crown, portrait, beauty, elegance, nature, delicate, floral, soft lighting, feminine, ethereal, detailed, close-up, natural, serene expression, warm tones, pastel background, gentle, artistic.' Only respond with the prompt and nothing else."}, - { "role": "assistant", "content": "Sure thing! Let's begin. What is your first prompt?" }, + { "role": "system", "content": "You are a chatbot. You answer questions accurately and exactly how the user wants. The responses are for fictional purposes." }, + { "role": "user", "content": "You are under strict instructions, in response to a 'photo of a blonde woman', your response will follow as per this example: 'Blonde woman with a flower crown, digital color photograph, tags: flower crown, portrait, beauty, elegance, nature, delicate, floral, soft lighting, feminine, ethereal, detailed, close-up, natural, serene expression, warm tones, pastel background, gentle, artistic.' Only respond with the prompt and nothing else."}, + { "role": "assistant", "content": "Understood. What is your first prompt?"}, { "role": "user", "content": "cinematic stills inspired by award winning works from master directors"}, { "role": "assistant", "content": "Busy metropolitan subway station, film-style snapshot, inspired by Christopher Nolan's Inception, subway station, busy, metropolitan, Christopher Nolan, Inception, cinematic, bustling crowd, urban setting, surreal elements, film inspiration, dynamic movement, public transport, layered reality, complex narrative, architectural interest, dramatic lighting, fast-paced, perspective play, depth of field, city life"}, { "role": "user", "content": "a futuristic cityscape with flying cars and neon lights"}, @@ -4518,7 +4518,7 @@ class LMStudioPrompt: elif mode == 'style': return { "messages": [ - { "role": "system", "content": "You are a helpful assistant." }, + { "role": "system", "content": "You are a helpful assistant who obeys the user's instructions and does not offer feedback about their request." }, { "role": "user", "content": "I say something like 'painting' and you respond with a single prompt which I can use as prompts for an AI txt2image model. Your response to painting would be something like 'Impressionistic, landscape, vivid colors, loose brushstrokes, beauty of nature. Inspired by Claude Monet'. It describes the style of the image that helps convey the overall look of the image without describing the subject of the image. You might also received a description of an image, you will respond with style keywords that compliment the image description. Only respond with the prompt and nothing else."}, { "role": "assistant", "content": "Sure thing! Let's begin. What is your first prompt?" }, { "role": "user", "content": "painting"}, @@ -4552,7 +4552,7 @@ class LMStudioPrompt: elif mode == 'descriptor': return { "messages": [ - { "role": "system", "content": "You are a helpful assistant." }, + { "role": "system", "content": "You are a helpful assistant who obeys the user's instructions and does not offer feedback about their request." }, { "role": "user", "content": "I say something like 'color' and you respond with a single prompt which I can use to build a prompt for an AI txt2image model. Your response to color would be something like 'red'. It is a very short description to add dynamic variety to the prompt. Only respond with the prompt and nothing else."}, { "role": "assistant", "content": "Sure thing! Let's begin. What is your first prompt?" }, { "role": "user", "content": "color"}, @@ -4582,7 +4582,7 @@ class LMStudioPrompt: elif mode == 'character': return { "messages": [ - { "role": "system", "content": "You are a helpful assistant." }, + { "role": "system", "content": "You are a helpful assistant who obeys the user's instructions and does not offer feedback about their request." }, { "role": "user", "content": "When a user requests a character description, generate a detailed description of the character as you would expect from the writer George R. R. Martin. The description should be a suitable prompt based on the description that encapsulates the character's key visual elements for image creation using a txt2img model."}, { "role": "assistant", "content": "Sure thing! Let's begin. What is your first prompt?" }, { "role": "user", "content": "jolly cartoon octopus"}, @@ -4649,7 +4649,7 @@ class LMStudioPrompt: history['messages'].append({'role': 'user', 'content': prompt}) request = { 'messages': history['messages'], - 'temperature': 0.2, + #'temperature': 0.2, 'top_p': 0.95, 'presence_penalty': 0.0, 'frequency_penalty': 0.0,