Bound by Blue
His kryptonite isn't green...
Fine Motions - Emotions - Long Video Testing for MiniMax H3
Touching, caressing and general emotions testing, mostly for me, I already know this model does the full range of motion and emotions. But, I need to know what words and prompts elicit which actions and emotion. All of this worked out fine, learned model prefers a full emotional state with body language rather than just an emotion.
Had issues with the lighting/colours: Latent encode/decode with latent upscale slightly darkens videos, reusing dark video to make next scene will darken the scene further, after 3 cycles of this the video's contrast is fucked, colours/lighting are all over the place for this video, tried to fix near the end, will find a better/permanent solution.
Voices are seed dependent? Voices aren't image driven, they're randomised around descriptions, giving "similar" voices for each section, I didn't know this, will need reinforcing with proper voice references. Also the double lip-sync for a voice line is a problem, I tried prompting it out but it will not fuck off, happens mostly when faces are too close together and character mouths are hard to define/separate, needs higher resolution testing but those tests eat time.
Still need a detailer/upscaler of some sort...considering using LTX2.5's spatial upscaling, it won't work great maintain characters and scenes though.
This video model still eats too much time to make long videos, not sure I can fix this, this video length is roughly my current time limit.
Pretty sure I fucked this up massively, but I learned everything I needed to so I guess I can't be mad...
Fine Motions - Emotions - Long Video Testing for MiniMax H3
Touching, caressing and general emotions testing, mostly for me, I already know this model does the full range of motion and emotions. But, I need to know what words and prompts elicit which actions and emotion. All of this worked out fine, learned model prefers a full emotional state with body language rather than just an emotion.
Had issues with the lighting/colours: Latent encode/decode with latent upscale slightly darkens videos, reusing dark video to make next scene will darken the scene further, after 3 cycles of this the video's contrast is fucked, colours/lighting are all over the place for this video, tried to fix near the end, will find a better/permanent solution.
Voices are seed dependent? Voices aren't image driven, they're randomised around descriptions, giving "similar" voices for each section, I didn't know this, will need reinforcing with proper voice references. Also the double lip-sync for a voice line is a problem, I tried prompting it out but it will not fuck off, happens mostly when faces are too close together and character mouths are hard to define/separate, needs higher resolution testing but those tests eat time.
Still need a detailer/upscaler of some sort...considering using LTX2.5's spatial upscaling, it won't work great maintain characters and scenes though.
This video model still eats too much time to make long videos, not sure I can fix this, this video length is roughly my current time limit.
Pretty sure I fucked this up massively, but I learned everything I needed to so I guess I can't be mad...






