as the get_images() function involves a `while True` loop while waiting for a response from ComfyUI and is not async, when image generation is running the entire UI becomes unresponsive for all users.
furthermore, when image generation takes too long, the Docker health check starts failing.
this is certainly a bad fix as it does not convert everything to async, but rather just puts the blocking loop in a separate thread. however, it works and it at least fixes the problem for now.
this commit adds three environment variables:
- COMFYUI_FLUX: determines whether Flux is used, the workflow is completely different so this is necessary.
- COMFYUI_FLUX_WEIGHT_DTYPE: sets the weight precision for Flux. you will probably want to set this to "fp8_e4m3fn" as the fp16 weights take up about 24GB of VRAM. optional, defaults to "default".
- COMFYUI_FLUX_FP8_CLIP: Flux requires two CLIP models downloaded, one of which is available in fp8 and fp16. set to true if you are using the fp8 CLIP weights.
this commit adds four environment variables:
- COMFYUI_CFG_SCALE
- COMFYUI_SAMPLER
- COMFYUI_SCHEDULER
- COMFYUI_SD3 (merely setting this at all will enable SD3 mode)
