This model does not currently work with native ComfyUI INT8, and will need to be converted with this tool.

Unless you already downloaded this model, I recommend finding a more recent INT8 quant from Silveroxides to avoid having to do this. converting this model actually.

Don't ask me why but the Silveroxides int8 convrot simple is performing worse than my converted version.

Evidence from my measurements:

8 Samples, 512x512, 20 steps each:

Metric INT8Bob INT8Silver
MSE ↓ 0.21857 ±0.01951 0.43413 ±0.05470
MAE ↓ 0.27689 ±0.01402 0.42484 ±0.02914
Max err ↓ 4.60856 ±0.16847 5.71161 ±0.20756
Rel-RMSE ↓ 0.42217 ±0.02033 0.61513 ±0.04038
SNR dB ↑ 9.10 ±0.46 4.95 ±0.56

16 Samples, 256x256, 20 steps each:

Metric INT8Johnson INT8_Silver
MSE ↓ 0.37593 ±0.04121 0.59410 ±0.04671
MAE ↓ 0.39571 ±0.02467 0.52660 ±0.02362
Max err ↓ 4.65401 ±0.19794 5.48461 ±0.16060
Rel-RMSE ↓ 0.54103 ±0.03184 0.70424 ±0.02955
SNR dB ↑ 6.44 ±0.55 3.53 ±0.38

This is an unavoidable double quantization due to the release state of Ideogram4.

The FP8 weights were cast to FP32 with the FP8 scales, then downcast to BF16 before being converted to INT8.

For use in ComfyUI with https://github.com/BobJohnson24/ComfyUI-INT8-Fast

Speed is 1.78x faster(2.03s/it) than FP8(3.62s/it) on my 3090, without compile.

~2x faster with torch compile.

After further inspection, it appears there may be quality issues with torch compiling this model.

Quick comparison:

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support