Little confuse with what loras to use? whats the imapct of them?
Little confuse with what loras to use? whats the imapct of them?
is there any one who can put some suggestion or explain how the workflow should use these loras?
For my specific cases I use the reference model: minimax_h3_hybrid_fl2va_ref2va_b25-49-int8 + qwen3vl_32b_minimax_h3_int8_convrot.
It should work similarly using the standard ref2va as well.
I test many of the turbo loras and to be frank, none of them work at 4 steps or less to produce anything workable.
After testing PDMD, Lightx2v, DMAD and LarryVH, TAOmate I have found that after testing them all only 2 of them stand out.
All used at 0.9mp 1280x736 8 steps with er_sde + sggm_uniform settings):
2.minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16 (The OG, good but suffers with fast motion too much)
https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
1.minimax_h3_ref2v_turbo_DMAD_4step_full_lora_avg_rank_39_bf16 (a good balance of everything actually workable and follows dialogue well)
https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras
Every other lora (like both Taomate, PDMD variants) have problems in my high quality test video, some have bad prompt following, weird cuts, or not following the dialogue exactly, some have too much jitter in teeth and trees and some deal bad with fast motion beyond accept. Toamate was adding random subtitles into the video which is a no no.
In theory they say PDMD is supposed to be the best one but not for me.
That said this is my use test case example, it may be different for you but those are my results.
speed is not everything. this vae produces blurred skin = fail.
speed is not everything. this vae produces blurred skin = fail.
Fast vae should only be used for previewing