FLUXinator (STD/GGUF), PuLID, Redux, img2prompt, controlnet, F1depth/canny, Detailer, SUPIR版本v1.5 (ID: 1255068)

ComfyUI Workflow for Single Image Generation

• needed custom node: RvTools v2

  • needs to be installed manually -> How to manually Install Custom Nodes.

• Links & Resources

• Workflow Video (iGEN)

• if you don't see a preview in the samplers, open the manager, in Preview Method choose: Latent2RGB (fast)

• Wildcards -> https://civitai.com/models/20868?modelVersionId=532723

if you run into issues with PuLID (tensor in different devices error) you can fix this by editing the pulidflux.py file like shown here:

https://github.com/Rvage0815/ComfyUI-PuLID-Flux-Enhanced/commit/0ed8d4abd40732040e88c29a4cde3383ee34a9b0

or delete the custom node and clone the forked version that already contains the fix from here:

https://github.com/Rvage0815/ComfyUI-PuLID-Flux-Enhanced.git

with 16 gb vram or less i would recommend to leave the purge nodes as they are. the startup time until a render starts is maybe longer, but in my tests without the nodes the vram was filled up to 99% after 1 or 2 canceled queues and then nothing worked anymore except restarting comfyui. i would also suggest to use the "unload models" button after chanceling a queue that helps a lot, otherwise it's possible that comfy tries to load flux twice into the memory. i can't say if the nodes are needed with a 4090, but i would bypass them first for testing and if it works without them they can simply be removed. (some are hidden behind other nodes you have to collapse them)

描述:

• rebuilded and fixed (e.g. sampling restarts when nothing has changed (seed, prompt etc.))

• removed nf4 support (doesn't work with loras so...)

• it's also possible to use the load image group, bypass the sampler and do the upscale/detailing on the loaded image

训练词语: