Repackages of Fal's Ideogram4-Fast and Ideogram4-Instant to ComfyUI compatible files, with int8 and nvfp4 conversion done with Silveroxides/convert_to_quant, targeting only attention and MLP tensors.
Officially recommended parameters:
- for Fast: 20 steps, CFG 1
- for Instant: 8 steps, CFG 1
Reported gen speed on a 5090 at 1536x1536:
- bf16: 0.85 it/s
- int8: 1.7 it/s
- nvfp4: 1.9 it/s
Note that Fast has QAD (Quantization Aware Distillation) so nvfp4 can be better than its larger variants, especially on long texts. Fal's readme of Instant says it doesn't have QAD but nvfp4 is just as good as the larger ones too.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support