Looking for a way to accept 2k and 4k footages as source and generate output in 2k and 4k?
as the inference is done in 1k or less at most. is there a way to keep all the resolution so all pipeline keep details.
or is there a detailer addon from the output of 720p to 4k, where ML regenerates fine texture and details in 4k after video.
LTX 2.3 updated few things, which preserve better details during the edit process. by good integration of image and video model, hope quality can bump up with few minor optimization.
Scene sometime loose tracking , and slides which does not make thing useable. is there a way to keep subject and bg track locked when there are panning , zooming, dolly, quick off and on camera blur etc. that would make this production ready.
FEATURE REQUEST: ABILITY TO COMBINE MULTIPLE TEMP VFX PLATES and place it in a XY location using some tool (using comfy). If a composited image is already given , then ability to insert multiple performance's high quality actors clip to be placed in the certain space according to reference.
long number of inference frames, where upwards of 3-5 minutes of footages can be injected , and background replaced in high quality. VACE no limit live streaming was done by someone, so i guess its just need batches of memory management , but I guess that would bring video Edit from few second of play toy to a serious consideration tool.
the mask and the edges flickers when isolating subject, so if external rough black and white mattes if can be added that would be good. please solve that edge issues. QwenVL is used under the hood , then relighting dont remove earlier shadows. is that a image edit problem or video issue i dont know, do keep an eye on that. just letting you know.
if camera pans , the reference image dont have beyond its frame. on that case 360 degree images can really help solve the hallucination, the best way would be to turn an image into an World map and let move the world camera to pan, zoom, move in 3d space. it would solve the world hallucination problem forever.
Looking for a way to accept 2k and 4k footages as source and generate output in 2k and 4k?
as the inference is done in 1k or less at most. is there a way to keep all the resolution so all pipeline keep details.
or is there a detailer addon from the output of 720p to 4k, where ML regenerates fine texture and details in 4k after video.
LTX 2.3 updated few things, which preserve better details during the edit process. by good integration of image and video model, hope quality can bump up with few minor optimization.
Scene sometime loose tracking , and slides which does not make thing useable. is there a way to keep subject and bg track locked when there are panning , zooming, dolly, quick off and on camera blur etc. that would make this production ready.
FEATURE REQUEST: ABILITY TO COMBINE MULTIPLE TEMP VFX PLATES and place it in a XY location using some tool (using comfy). If a composited image is already given , then ability to insert multiple performance's high quality actors clip to be placed in the certain space according to reference.
long number of inference frames, where upwards of 3-5 minutes of footages can be injected , and background replaced in high quality. VACE no limit live streaming was done by someone, so i guess its just need batches of memory management , but I guess that would bring video Edit from few second of play toy to a serious consideration tool.
the mask and the edges flickers when isolating subject, so if external rough black and white mattes if can be added that would be good. please solve that edge issues. QwenVL is used under the hood , then relighting dont remove earlier shadows. is that a image edit problem or video issue i dont know, do keep an eye on that. just letting you know.
if camera pans , the reference image dont have beyond its frame. on that case 360 degree images can really help solve the hallucination, the best way would be to turn an image into an World map and let move the world camera to pan, zoom, move in 3d space. it would solve the world hallucination problem forever.