- Can be used via online services or within Comfy.
- Pro version of Flux only via online partners
- Schnell & Dev available at home.
- Images are high quality, midjourney level, but with a fairly narrow range out of the box.
- All people tend to look like models
- Photography looks very high end with shallow depth of field.
- Less flexible than SDXL at more stylized results.
- Matteo’s commentary on his deep dive video is great.
- Responds very well to LoRA training. Likeness, concepts, even text working well with small amount of images.
- Very resolution dependent - 1024x1024 and similar 1k resolutions work the best. (note: I feel like I’ve seen a list of res combos that works well for it?)
- As of August 2024, controlnets exist and they’re okay. IPAdapter exists and quality is bad, as per Matteo.
- It’s helpful to have 24+ gigs of VRAM.
Comfy basics, models, installation: https://comfyanonymous.github.io/ComfyUI_examples/flux/