Ideogram V4 Fast and Instant: Speed-Distilled Variants by Fal for ComfyUI
Fal releases Ideogram V4 Fast (20-step FP4, 5s inference) and Ideogram V4 Instant (8-step BF16, 2s inference), speed-distilled variants with no-CFG support. Available in int8-convrot format for ComfyUI via Hippotes.
Fal has released two speed-distilled variants of the Ideogram V4 text-to-image model — Ideogram V4 Fast and Ideogram V4 Instant — bringing significant inference speed improvements while maintaining visual quality.
Both variants are now available on Hugging Face as open-source weights, with ComfyUI-ready int8-convrot versions provided by the community.
What's New
The two new variants address different speed-quality tradeoffs:
| Variant | Steps | Precision | Inference Time (NVFP4) | Key Technique |
|---|---|---|---|---|
| Base (NVFP4) | 20+ CFG | NVFP4 | ~11s | Original |
| Fast | 20 | FP4/NVFP4 | ~5s | CFG distillation + QAD |
| Instant | 8 | BF16 | ~2s | No-CFG, 8-step distillation |
Both models are derived from ideogram-ai/ideogram-4-fp8 using quantization-aware distillation (QAD) and CFG distillation techniques.
- Ideogram V4 Fast (hf.co/fal/ideogram-v4-fast): An FP4-targeted, 20-step checkpoint that folds the guided prediction into a single conditional branch, eliminating the need for unconditional scoring. This combines speed distillation with NVFP4 quantization.
- Ideogram V4 Instant (hf.co/fal/ideogram-v4-instant): An 8-step checkpoint that operates without CFG (classifier-free guidance), further reducing computation. These are BF16 weights from before the QAD stage, offering maximum compatibility.
ComfyUI Support
Community member Hippotes has packaged both variants for ComfyUI at Hippotes/Ideogram4-Fal-ComfyUI, including int8-convrot versions:
- BF16 checkpoints for Fast and Instant
- int8-convrot quantized versions for improved performance on Nvidia GPUs (20/30/40/50 series)
With the no-CFG architecture of these variants, ComfyUI workflows no longer require the Dual Model CFG Guider or CFG Override nodes, and there is no need to load the unconditional diffusion model — simplifying the workflow setup.
Speed Benchmarks
Benchmarks shared by the community on an RTX 6000 PRO (Ada) GPU show:
- Base NVFP4: 11 seconds per generation
- Fast NVFP4: 5 seconds per generation (2.2x speedup)
- Instant NVFP4: 2 seconds per generation (5.5x speedup)
Technical Background
In a companion blog post, fal detailed their approach to achieving sub-second Ideogram V4 inference. By combining quantization-aware distillation, CFG distillation, and FP4 targeting, fal reports a 6x speedup from the original 2.75s (on their infrastructure) down to 0.44s with no visible quality loss.
Availability
- Ideogram V4 Fast: Hugging Face — FP4, 20-step, CFG-distilled
- Ideogram V4 Instant: Hugging Face — BF16, 8-step, no-CFG
- ComfyUI versions (int8-convrot + BF16): Hippotes/Ideogram4-Fal-ComfyUI
Comments
Sign in with GitHub to join the discussion.