Saw a thread on Reddit about MiaoshouAI Tagger and tried their work flow, this is based on that so credit for the core goes to them, here's a link to it: https://raw.githubusercontent.com/miaoshouai/ComfyUI-Miaoshouai-Tagger/refs/heads/main/examples/miaoshouai_tagger_flux_hyper_lora_caption_switch_workflow.png

The core operates by dropping your image in the load and resize node and set the toggle for if you want to resize it or not, the green node next to the Note on the left hand side has a true/false toggle. Set it to true for Img2Img or false for Txt2Img, there are 2 float nodes on the right hand side of that to set Denoise for both nodes. The captioner uses a LLM that auto downloads and is a fine tune of Florence2. It creates a natural language T5 prompt and a tag style clip l prompt.

Of course had to tinker and add a few things:

Force Clip so that can be offloaded to CPU, takes longer for the Clip Text encode, but with the size of Flux Dev being 23GB it's worth keeping the VRAM free for the model. Same goes for smaller models on lower VRAM cards. Also added Multiply Clip, ModelSamplingFlux and Detail Daemon nodes to the main part. I do have it set up for the Verus Vision model as I prefer the skin textures and think it looks more natural compared to Flux, as it's a fine tune of a de-distilled version the recommended CFG is 3.5. If using a standard distilled version of Flux, change that to 1. Detail Daemon is also a lot more sensitive for non distilled models, you can increase the settings on there if using standard Flux.

I also added 3 groups. ControlNet - I'm using the Flux ControlNet Union Pro model and Depth Anywhere V2 as I find them to be pretty effective with Flux. The second group is PuLID to use a face reference and the last group is UltimateSD Upscaler. These all have bypass switches and you can turn off any of the groups you don't want to use.

The faces could do with some detailing afterwards, but it's fun to play with.