I2V! I have used accidently used the ref2V turbo lora in combo with fl2V model, 8 steps and the character preserving and face consistency is 10/10!!
I have some extreme proportion image, generations with no lora are losing the connection to the image character within the 1st steps same with the usual turbo lora combos. the prompt line "\nPreserve the identity of <Girl 1>, her exact face, facial features, eye shape, nose, lips, skin tone, hairstyle, clothing and body seems to have no effect. Thx for the lora lightx2v-team!!
Interesting find.
Please share your workflow.
What lora exactly. what model loader, what settings. etc.
R2V is omni , which means you can you it as first frame also with better stability and consistency.. i dont know why one still use F2V.. with R2v you can also get away with alot less steps
Interesting find.
Please share your workflow.
What lora exactly. what model loader, what settings. etc.
Still there is only 1 lora minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors. I am using it in the standard worklow with the comfy kitchen node. 8 steps. models I tried in combo: minimax_h3_fl2va_pruned_bf16.safetensors, minimax_h3_fl2va_int8_convrot.safetensors. TE n4 or int8 convrot
hmm I tried minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors for fl2va, while motion is good, the image is blurry compared to minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors has slow motion though
Interesting, I used the FL2V LoRA on the Ref2V Model and got excellet identity preservation.
If anyone interested - try using ref conditioning node on fl2va model, 2 steps ref2va turbo and 4 steps with fl2v 768 or v4_step600_ema. And you can even add ref_patch above it.