Just a quick benchmark
https://github.com/kijai/ComfyUI-KJNodes (Diffusion Model Loader KJ)
https://github.com/kijai/ComfyUI-SolAttn_triton (Patch Sol-Attn)
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3 (Spectrum Apply MiniMax H3)
Whenever I try to use the spectrum node it makes my generations go twice as long, not sure what I'm doing wrong.
I find it hard to fit the workflow into my 4090 24gb device. Would you mind sharing your workflow setup for reference? Thanks
Thank you Unt! The patch sol attn node is a good 15-20% boost. I had never used it before!
From my experience you get to pick between Spectrum and EasyCache. If you use one, you can't use the other. EasyCache is a bigger speedup on anything over about 100 frames in my testing. Spectrum did shave off about 15% in my testing. But EasyCache is closer to 25%. Spectrum also has some effect on motion I think. I tested the provided Spectrum settings and that matched my previous default test in terms of speed. KJ's sage attention patch is a must.
I've also had much better results using LCM and Beta instead of the default res_multi and simple. But no speed up.
Tried LCM and Beta - somehow the sound was totally off, only weird ringing noises.
Why not MiniMax H3 Mem Eff Sage Attention Patch (from the latest KJNodes) instead of the generic KJ Sage Patch ?
We are having too many nodes from different threats, I'm getting dizzy - easycache, ComfyUI-MiniMaxH3-Cache, spectrum, sol, three ways of enabling sage, different sampler+scheduler combos, oh boy :D
Do we just replace the Patch sage attention KJ with the H3 Mem Eff Sage Attention?
Kitchen Attention on an RTX 5090: no measurable speed difference for H3
Since attention speedups keep coming up here, a datapoint from a 5090 (32 GB, ComfyUI v0.33.0 nightly, PyTorch 2.10+cu130, minimax_h3_fl2va_pruned_int8_convrot + nvfp4 text encoder):
ModelAttentionBackend → "comfy kitchen attention" vs default pytorch attention, identical fixed graphs, warm model, alternating arms:
- 4-step turbo LoRA, 1344×768, 56 frames: 16.17 s both arms (3 seeds each, spread <0.2 s)
- 20 steps, same shape: 48.33 s vs 48.32 s
So the ~20–30% gain reported on a 3060 doesn't transfer to Blackwell — consistent with the note that 50-series cards "may not see much difference." Harmless, but no reason to add the node on a 5090.
Benchmarking tip: re-submitting a bit-identical graph hits ComfyUI's execution cache and returns instantly — vary the seed per run.