just test with 3x3 grid various pose of consistent face with flux,flux fill,lora, eyes detail,chin detail,etc..

write a python script to run 4 8G vram 3060Ti in parallel with wan2.1 480P i2v model,each output video is 73 frames in total,then combine all of'em with ffmpeg

deepseek R1 prompt instruction=>flux image generation=>reactor face model swap=>consistent scene and character 2*2 grid image run in local,then run multiple comfyui instances on each 3060 ti 8G vram simultatenously thr wan i2v 480P model online server,combine all 4 videos at last

if u have multiple workflows like consistent character images ,faceswap with reactor,and wanna to run on comfyUI API server,u should share and cache the diffusion model and clip model to avoid reloading of'em each time,otherwise the server will be overwhelmed and in disconnection

if u upgrade cuda from 12.4 to 12.8,you have to recompile sageattention for usage,the inference speed will be doubled compared with sdpa,the latter will be will around 20 mins than the previous around 13 mins in case of the wan2.1 FLF model

wan 2.1 ATI I2V model testing

Wan2.1 FusionX the LoRA model generates video faster,81 frames with 8 steps took around 160s,161 frames with 8 steps took over 4 mins

wan fusionX VACE test

wan FusionX_Ingredients T2x workflow video game scene testing which include the assassin's creed,black myth wukong and pirate of the caribbean

wan fusionx i2v wrapper ingredient test

multitalk text to speech which built on eSpeak NG TTS synthesizer