Video - https://youtu.be/T7nufrDmc9A
DeepSeek. It's all the rage at the moment, so I figured I'd give it a spin in ComfyUI to see how good it was at prompt enhancing for image generation!
The following models were used to compare against:
DeepSeek-R1 (14B)
Gemma2 (9B)
Llama 3.2 (3B)
Olmo (7B)
Phi4 (14B)
The TL;DR is yes, it's very good. Is it the best? Not fully convinced yet :) I'm liking Olmo at the moment, but time will tell.
I also found the new gradient estimation sampler gave decent results with Flux, so certainly taking a look at if you've not done so as yet. It certainly seems to pair nicely with the kl_optimal scheduler.
Links
https://ollama.com/library/deepseek-r1
Workflow
Just a single workflow this time as it's simply for testing out the prompt capabilities of various large language models.
It includes various extra nodes from previous videos, such as Detail Daemon and WaveSpeed.
If you don't have ollama installed yet, you'll need to do that first, as well as download whatever models you want to test from the ollama library, E.g.
ollama pull deepseek-r1:14b
LLM mode can be turned off too, should you just want the basic prompt.
The regex node used is from comfyui-lf - https://github.com/lucafoscili/comfyui-lf (as apparently manager can't detect this one)
WaveSpeed (covered previously) - https://github.com/chengzeyi/Comfy-WaveSpeed
Derfuu - https://github.com/Derfuu/Derfuu_ComfyUI_ModdedNodes
Example screenshot of a completed workflow run:
Also included is the full-size montage image of the test results.