Getting weird video artifact on ref2va
Anyone know why? I changed seed length, LoRA, models, steps, scheduler, resolution, but still get the same flick exactly at frame 51 here 😀
Bypassing the VEDA node, and it works. This is ref2va workflow with ref image as the first frame
Can you share more settings? We have noticed similar effect from using double samping (selflit or some kind). At low resolution, the token/tiles are so few that it bearly hold enough infromation. We are working on fixing it.
It's just a basic ref2va workflow. I think using the ref image as the starting frame causes it; but I'm not sure. I haven't tested a lot.
i change settings around, and it still has the artifact at that exact frame
Really impressive results thanks for your work Team Veda.
I also have had these white flashes sporadically. Seems to happen at shot changes most commonly. Haven't noticed if related to camera direction.
Main model: minimax_h3_ref2va_pruned_int8_convrot.safetensors
Turbo lora from lightx2v minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors
Maybe getting better results since using their TE instead of an abliterated one: qwen3vl_32b_minimax_h3_int8_convrot.safetensors
Doing generations that align perfectly with the Veda runs, the node reports no padding. 10 or 15 seconds (snapped to 17n5).
My anecdotal results in case useful.Still testing.
This might be a possible fix as reported in the Github discussion: https://github.com/veda-sparse/Veda-on-ComfyUI/issues/13
Original solution by Rkkss: IN full attention layer, set it to "0,1,49"
It worked for me!