Model Timing Comparisons for 5060ti 16gig VRAM
Coming over from Civit comments with workflows and videos.
Stock 20gig Int8 convrot, ran in 105s (after a warmup video) Comfy native w4a8 from your civit page, ran in 95s after warmup.Pretty similar results but no marked speed increase by running the model that should fit into my vram. Larger model has better quality (though I'd say only like 10%?)
Sorry, seeing this late. Thanks for the side-by-side — the numbers make sense, and there's likely a 2× sitting in your launch flags.
Why the smaller model isn't faster: both files you tested are comfy-native quantized formats (int8-convrot and w4a8). Whether they run fast depends on which comfy-kitchen backend is live on your box, not on file size. On a torch build older than cu130 (very common), the CUDA backend is disabled and the Triton one is off by default — so both formats fall back to "eager", which unpacks the weights to bf16 every step. That costs the same either way, which is why 105 s ≈ 95 s. The w4a8's win is memory headroom (longer clips, higher res on 16 GB), not seconds per step — until the backend is enabled.
Check it in one look: in your ComfyUI startup log, find the lines starting Found comfy_kitchen backend. If cuda and triton both say 'disabled': True and only eager is enabled, that's it.
The fix: add --enable-triton-backend to your ComfyUI launch line (Triton has to be installed — on Windows that's the triton-windows package). Measured on a 24 GB card last night, same seed and settings: 36.7 → 17.6 s/it, −52% for a comfy-native quant file. GGUF files don't go through comfy-kitchen at all, so they're unaffected either way — which is why they felt "just as fast."
Quality: your read matches ours — the larger int8 is a little cleaner. w4a8 costs a touch of texture, worth it when you need the room.
The next pack version prints a warning at model load when only eager is live, so nobody pays 2× by accident.
So, mixed bag on this. I have cuda enabled, but triton disabled.