Very bad prompt adherence
I can write stuff and the prompt is completely ignored. 99% of videos are made as a tv ad which is stupid. I think that was main bulk of training data.
Dumb example: I can write, human looking at flower then looks at a cat. The video will NOT BE like that. I can write detailed prompt with times - doesnt matter. There is no setting that makes to adhere to prompt? no cfg? anyone?
Minimax has the best prompt adherence from all open source models I've used. Are you following official prompt guidelines? Basically you've got two ways. First is to process your prompt through api using Context-IR as they mentioned it's crucial for best output quality. There's also community solutions which tries to reverse-engineer Context-IR final output but right now all solutions are so so we just gotta wait ecosystem is too fresh. Best whan you can do for free right now is just use good flagship llm, fill it with your idea and provide skills https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills
I never got bad prompt adherence following this method.
Minimax has the best prompt adherence from all open source models I've used. Are you following official prompt guidelines? Basically you've got two ways. First is to process your prompt through api using Context-IR as they mentioned it's crucial for best output quality. There's also community solutions which tries to reverse-engineer Context-IR final output but right now all solutions are so so we just gotta wait ecosystem is too fresh. Best whan you can do for free right now is just use good flagship llm, fill it with your idea and provide skills https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills
I never got bad prompt adherence following this method.
I definately was not using context-IR. Thanks, i will try your advice!!