I’ve been messing around with AI image generation for a while, and lately, I’ve been stacking nodes like Deepseek, Flux, and ComfyUI to see how far I can push the realism. The results? Honestly, better than I expected. Here’s how I got it running.
First, I grabbed the Deepseek Prompt Generator node from this GitHub repo. It’s a custom node that takes a basic idea and spits out a ridiculously detailed prompt—way better than anything I’d come up with manually.
Getting Deepseek Running
You can hook it up two ways:
- API: If you’ve got access to Deepseek’s API, just drop your key into the config.
- Local (Ollama): I went this route because I didn’t want to deal with API limits. Installed the 14B model (
ollama run deepseek-r1:14b) and it worked without much fuss.
ollama run deepseek-r1:14b Loading Up Flux
For the actual image generation, I used Flux v1.5 pruned—just dropped the checkpoint into models/checkpoints. Flux handles the base image, and then I layer on extras like the Ultra Real Photo LoRA (found here) for hyper-detailed skin and textures.
The Workflow
After importing the workflow into ComfyUI, it chains together:
- Deepseek Prompt Generator – Turns my half-baked ideas into pro-level prompts.
- Flux Model – Generates the initial image.
- Upscaler (4xNMKD) – Because who doesn’t want more pixels?
- Face/Hand Fix Nodes – Because AI still struggles with fingers, let’s be real.
Prompt Magic
Here’s the wild part: I fed it something simple like “woman in a café, blue eyes, New York” and it expanded it into this cinematic-ready paragraph with lighting, ambiance, and even wardrobe details. If the output feels off, I tweak it directly in the Write Your Prompt node, but most of the time, it nails it.
The upscaling and LoRAs take the images from “pretty good” to “wait, this isn’t a photo?” levels. Still testing, but so far, it’s my go-to for realism.
https://huggingface.co/identityailabs-com/ultra-realistic-lora/tree/main
