ref got transformed before animation even begun
ref image got transformed before animation even begun
-- i been testing with locked seed
-- only model that did not keep the look of the ref image
-- tested 5 different models --hybrid/default base ..and other fine tunes
interesting note --- i built the truck -- that the image is of --- it is 100 percent custom -- with a chrysler 300 grill on a ford ranger
--what is interesting is that this model used a older model ranger front end in replacement of my custom build on a ford ranger
this model
one of the others
--- what do you think happened to make that happen
minimax_h3_ref2v_turbo_4step_v0.1
I think this version generates it perfectly.
It seems like the other version isn’t working properly.
If I mix in another Lora, they don’t eat it well either.. It seems to work best only when you use just one Lora. minimax_h3_ref2v_turbo_4step_v0.1
minimax_h3_ref2v_turbo_4step_v0.1
I think this version generates it perfectly.
It seems like the other version isn’t working properly.
If I mix in another Lora, they don’t eat it well either.. It seems to work best only when you use just one Lora. minimax_h3_ref2v_turbo_4step_v0.1
okay cool. i will try without turbo as well, thanks for reply
ref image got transformed before animation even begun
-- i been testing with locked seed
-- only model that did not keep the look of the ref image
-- tested 5 different models --hybrid/default base ..and other fine tunes
interesting note --- i built the truck -- that the image is of --- it is 100 percent custom -- with a chrysler 300 grill on a ford ranger
--what is interesting is that this model used a older model ranger front end in replacement of my custom build on a ford ranger
this model
one of the others
--- what do you think happened to make that happen
Thanks for the detailed feedback! I haven't specifically stress-tested this model for your exact reference-to-animation setup.
To achieve better visual quality and dynamic motion, I tuned this model with a more aggressive noise diffusion strategy in the early inference stages compared to the official base model. However, this aggressive early-stage diffusion is likely what causes the initial frame to distort and overwrite your custom features before the animation kicks in.
In my own testing, I found that using the v0.1 version of the 4-step LoRA (at a strength of 0.75 to 1.0) helps smooth out this aggressive early noise phase. It gently guides the initial inference so the model has more headroom in the mid-to-late stages to refine details, preserve your reference image features, and deliver stronger overall performance. Give that a try and see if it helps!
okay cool.. i will make a few adjustments for next try 😁 thank you for reply