I've learned a lot from others, and I love tweaking ComfyUI to my needs. This is the workflow that works for me with Flux (dev, fp8). Key features :

• All the stuff I need in the same spot (source image with denoise for IMG2IMG, loras, batch size, picture resolution)

• I can prompt in my natural language (french) and in English (French prompt is Google translated and concatenated with the English one)

• Picture are saved with enough metadata for Civitai (did not manage to get Loras weight)

• CLIP is forced on CPU/Ram so I can fit Flux in my VRAM (12go), on my 4070 Super I get around 3s/iteration when generating 2 pictures, so about one minute if I stick to 20 steps

• Image presets for resolution (I hate typing a picture size in pixel, having a dropdown with the right resolutions is so much better. And a toggle to switch from landscape to vertical). Hint image get resized to this resolution when used as latent.

Some tips :

• Install missing nodes with ComfyUI Manager

• most nodes are collapsed, you can click the circle to expand them if you need to change something (like the source language in the translator node)

• to avoid spaghetti I use Get/Set nodes, Setter are hidden below their related nodes (ie: Set PROMPT_EN is below the English prompt)

• I use 2 variants of Clip_L, Long-ViT-L-14-BEST-GmP-smooth-ft.safetensors which supposedly is better for long prompts, and ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors when text matters. Differences are not huge, you can stick to Clip_L