VRAM
How much GPU Vram needed to run Trubo model for a HD image?
any minimum hardware spec?
based on the size of the weights of the diffusion model it should be somewhere around 16Gb of minimum VRAM =( less than it the only way is use a GGUF or FP8 version
just use wan2gp, i did a 1080p image in less than 2minutes using just 5vram
if you run it in things like comfyui, they handle the ram swapping intelligently. a 10gb-ish, even bigger, fp8 or gguf vrsion of most image models does ok, if relatively slow on my old 5gb vram laptop (4050).
MAybe not flux.2 dev (30gb) or hunyuan 3 (80gb).
I am running the RTX A2000 8GB GPU using the default model configs 1024x1024
I was getting OOM errors until I did the following.
I got 1024x1024 Image generation working on the 8 GB VRAM class RTX A2000 setup in NORMAL_VRAM with DynamicVRAM enabled.
Docker rebuild details:
CUDA base image change to 12.6.3 cudnn runtime
PyTorch wheel index change to cu126
Installed torch, torchvision, torchaudio versions
CUDA-specific runtime libraries pulled by torch:
cuDNN, cuBLAS, CUDA runtime/NVRTC, cuFFT, cuRAND, cuSOLVER, cuSPARSE, cuSPARSELT, NCCL, NVJITLINK, NVTX, cuFile
in comfyui with new dynamic vram feature i think you need just 24 gb ram and 8 gb vram minimum
3060 ti here, 8GB VRAM, fp8 quant gets about 3.9 seconds per iteration, and it generates images about as fast as the fp8 version of Flux2 Klein 9B can edit them, which on my machine is about 30 seconds, max. The speed is supremely tolerable in exchange for the image quality; photography-focused samplers like sa_solver can produce very clear, clean results. It's nice being able to run slightly larger models like this in Comfy now without them OOM-ing.
Is running this model comfortably even possible on AMD GPU?
On my 9060XT 16GB, it takes 200-300 seconds for one 1MP image?..
ComfyUI 0.27 Krea2 FP8