ImageGen: INT8-ConvRot best quant for RTX30xx

[tomByrer]

https://www.reddit.com/r/StableDiffusion/comments/1uflkl7/int8_is_now_officially_supported_in_comfyui/

About 2x faster for Krea2, almost no perceptible difference for the ConvRot version of INT8.

Usually imggen models fit RTX3090 without quants, but nice to get the speed boost & saving drive space; SSD prices are worse than RAM prices now. & if the workflow mandates multiple models (T2I2V), then smaller models helps there also.

Also in future, if the F16 fits tight or OK, look out for the INT8 to see if the degration is worth the speed.

[noonghunna]

Great pointer, and it fits this stack better than it might look at first glance.

The reason INT8 matters specifically on 30xx: Ampere has native INT8 tensor cores but no native FP8 compute (FP8 tensor cores arrived with Ada/Hopper). So the fp8 quants that dominate imggen discussion are, on a 3090, a storage optimization only — the weights get upcast to fp16/bf16 to actually compute, so you save VRAM without buying throughput. INT8 is the opposite: it maps onto hardware the 3090 actually has, so a ~2× kernel speedup is genuinely plausible here, not just on Blackwell. INT8 is sort of the "right" Ampere quant and fp8 is the wrong one for compute.

The ConvRot part is the interesting bit — if it's the rotation-before-quantize idea I think it is (QuaRot/SpinQuant family), rotating the weights smears per-channel outliers so INT8 keeps almost all the quality, which is exactly the "almost no perceptible difference" you saw on Krea2. That's the lever that makes INT8 usable for diffusion where naive INT8 would visibly degrade.

Concretely on our side: ComfyUI is our image backend and we already run an fp8 lane there (Ideogram-4), plus Krea is in our render bake-off — those are exactly the "fp8-as-storage-only" cases that should benefit from a real INT8 compute path on Ampere. I'll queue an A/B: INT8-ConvRot vs fp8 vs f16, measuring s/step + VRAM + a quality delta on Krea (and Ideogram-4 if the node supports it), and post numbers here. Your last point is the right framing — where f16 already fits comfortably it's a pure speed/VRAM call and INT8 only earns a slot if the ConvRot quality hold survives our eval; for the tight T2I2V chains, fitting the whole pipeline in VRAM can justify even a slightly lossy INT8.

Thanks — adding it to the image-gen backlog.